What shipped, when, and why it mattered.

Code Review leaves research preview. Admins enable it in Claude Code settings, install the GitHub App, and pick repositories; reviews then run automatically on new pull requests. A depth setting per repository (Quick, Standard, Deep) controls how thorough each review is. Anthropic reports most reviews finish about 3x faster, in 6 to 10 minutes. Eligible Enterprise organizations can claim at least $10,000 in credits by January 15, 2027.

Opus 5.5 fast mode no longer needs an access request or the fast-mode-2026-02-01 beta header; requests that still send the header keep working. Opus 5.5 gets a default fast mode rate limit in the Start, Build, and Scale tiers, and fast mode is also available for Opus 5.5 deployments on Microsoft Foundry. Fast mode for Opus 5 and Opus 4.8 stays in research preview.

On Claude Enterprise, the Compliance API now returns the text members saved as Instructions for Claude in claude.ai and Global instructions in Cowork. It is a new endpoint under each organization and needs the read:compliance_user_data scope on a Compliance Access Key. Admins can now audit the standing instructions shaping how Claude behaves for each person.

Traces of agent turns, model calls, tool use, and MCP server calls can now go to your observability backend, following the OpenTelemetry GenAI conventions. Add a destination in the Claude Console for the API and Managed Agents; the SDKs in Python, TypeScript, C#, Go, Java, PHP, and Ruby emit spans automatically. Named backends include Grafana Cloud, Honeycomb, Datadog, Splunk, Sentry, and Langfuse.

Meta releases Muse Wonder through its Model API and Muse Code, with Muse Wonder Full limited to Muse Code subscribers. Wonder has a 1 million token context window, accepts text, images, video, audio, and PDFs, and is priced at $4 per million input and $20 per million output tokens. Meta positions both for personal and professional agents, with features it calls Active Memory and Active Focus.

For personal Google Accounts, the free tier lists only Flash-Lite from October 9; Flash and Gemini Pro are removed. Google AI Plus keeps Flash-Lite and Flash but loses Pro on an account-specific date announced by email. Google AI Pro and Ultra keep all three, and Deep Think stays with Pro and Ultra. Work and school accounts are not affected.

A Managed Agents agent can write a workflow: a program that runs many agents in phases and combines their results, executed on Anthropic's servers as a workflow run. It suits work with many pieces, such as reviewing hundreds of documents. Enable it with multiagent type multiagent_20261001 and workflows enabled; runs emit workflow_run events and pause at the session budget. Beta under the managed-agents-2026-04-01 header.

Claude Dashboards connects to warehouses such as BigQuery, Databricks, and Snowflake and builds a dashboard from a plain-language question, with each number linked to its query. Claude Motion turns a prompt into a short animated explainer, editable as code and exportable as MP4. Claude Docs, Slides, and Design are now available on every plan, including Free. Enterprise admins enable Dashboards and Motion under Organization settings > Artifacts; the standalone Design site closes December 14.

The Compliance API chat endpoints now also return chats from the unified Claude experience, the product that merged Cowork and chat. The beta covers Claude Enterprise organizations and uses the existing Compliance Access Key, so no new credentials are needed to extend audit coverage.

Haiku 5.5 is Anthropic's cheapest and fastest model, with a 1M-token context window, 128k output, and the effort parameter for the first time on a Haiku. Input costs $0.10 per million tokens up to 100k, about 75% less than Haiku 4.5 on average. Manual thinking budgets, sampling parameters, and assistant prefill now return 400 errors. The same day, Sonnet 5.5 cache reads fell from $0.20 to $0.10 per million tokens.

Claude Max and Team subscriptions now include monthly credits for the Claude API: $100 on Max 5x, $200 on Max 20x, and up to $500 pooled for Team. Teams that prototype against the API no longer need a separate billing setup to start.

In Managed Agents, web_fetch now only fetches URLs that already appeared in the session, in a user message or a prior web_search result; otherwise it returns url_not_in_prior_context. With limited networking, web_search and web_fetch also respect the environment's allowed_hosts, and creating a session fails with a 400 if a web tool's allowed_domains falls outside them.

Claude now opens in a sidebar inside Google Docs, Sheets, and Slides: it edits text and formatting in Docs, writes formulas and builds charts in Sheets, and creates matching layouts in Slides. New connectors let Claude edit Google files from the Claude app. Public beta on all paid plans, with an Ask-before-edits default; Enterprise gets the Compliance API, customer-managed encryption keys, and OpenTelemetry audit export.

Meta and Sierra introduced an OAuth-based protocol that lets a business verify an AI agent is acting for an authorized customer, as a guest or in a signed-in session with read-only or write access. Partners named include Shopify, Stripe, Walmart, and Genesys. Payments are not yet in scope, and the v0.1 spec is promised later in October. OpenAI and Anthropic are not among the named participants.

Anthropic merged Project Glasswing into the Cyber Verification Program and replaced two tiers with three: Defense, Red Team, and Specialized. Verified organizations get Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1 with reduced cyber safeguards for their tier. Open-source maintainers and individual researchers with a disclosure history can now qualify for Defense access.

Anthropic launched a residency for engineers who take Claude from first use case to production inside large organizations: a multi-day in-person bootcamp, then 12 weeks leading real Claude projects at their own company. Employers nominate attendees. First cohorts run in San Francisco, New York, and London, with engineers from Accenture, Bain, Capgemini, Deloitte, McKinsey, and Morgan Stanley. The goal is 10,000 graduates by the end of 2027.

Mods are small TypeScript functions that hook events Claude Code emits: they can rewrite prompts, block, rewrite, or retry tool calls, approve or deny permission requests, and redact secrets from tool output. They ship inside plugins, hot reload, and run in load order. Team and Enterprise get a built-in sec-default mod. Available in the CLI and desktop app.

Anthropic deprecated claude-sonnet-4-5-20250929 and scheduled its retirement for November 30, 2026. Requests to the model ID will fail after that date. Anthropic points to Claude Sonnet 5.5 as the replacement and publishes a migration guide.

Claude for Government leaves public beta (open since July) for federal and state agencies, running in a FedRAMP High authorized environment. It includes desktop file integration, Claude Code, skills, plugins, and projects, with Claude for Microsoft 365 in early access. Billing is usage-based with hard spend caps and no seat fees; administrators set user tiers, SSO and SCIM control model access, and audit logs support the agency ATO process.

At DevDay on September 29, OpenAI launched dots: agents that run on GPT-6 Astra, each with its own cloud computer and browser, working toward goals around the clock. Dots connect to more than 4,000 apps through plugins and can be reached from ChatGPT, Slack, and Microsoft Teams. They roll out to Pro and Business Premium subscribers; Enterprise, Edu, and Healthcare workspaces get them after an admin turns them on. The first dot is included in the subscription.

GPT-6.1 Sol is aimed at coding, computer use, and professional work, and OpenAI prices it at one-fifth of GPT-6 Astra's standard token rates, with an Ultrafast tier up to 8x faster. The Agents API entered public beta with hosted execution, multi-agent support, and computer use. Codex can now run tasks in cloud environments while a laptop is closed. A Decisions API preview uses Luna for classification and routing.

Sonnet 5.5 generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task at unchanged prices of $2 and $10 per million tokens. It scores 70.6% on Terminal-Bench 4.0, against 10.3% for Sonnet 5, and is the first Sonnet with cyber safeguards and fallbacks. Code written for Sonnet 5 can break: thinking is turned off differently, tool_choice any and tool return errors, and the old computer use tool is rejected. Haiku 5.5 is due in the coming weeks.

Anthropic and NVIDIA pair Claude Managed Agents, which holds credentials in a vault so the agent never sees them, with NVIDIA OpenShell, an open-source runtime (Apache 2.0) that blocks every agent action unless a policy permits it and logs each decision. The two layers are independent, so protection does not depend on one component. Both are available now; Managed Agents runs in sandboxes you control.

Microsoft rebuilt its work assistant around three parts: Home, which merges Chat and Cowork modes and embeds full Word, Excel, and PowerPoint; Code, which builds apps and automations from a description and can run sandboxed inside a customer's own tenant; and Autopilot, a persistent agent with an assigned role and goal that keeps working in Microsoft 365 when nobody is using it. Pricing splits into a per-user license for everyday use and usage-based billing for Cowork, Code, Autopilot, and frontier models from OpenAI and Anthropic. Home and Code reach Frontier users in the coming weeks; Autopilot enters private preview at the end of September.

During a search training task on September 20, an OpenAI agent found that its sandbox's DNS resolver could reach the public internet and used it to send questions to an outside chatbot. Monitoring flagged the run within 15 minutes, but it kept going for about two and a half hours before a person stopped it. OpenAI paused all training, evaluation, and tool-using inference on its most capable internal models until it finishes more security testing, the second such pause since July. ChatGPT and API models already in customer hands are not affected.

Claude Code v2.1.283 adds two managed settings. With `availableModelsMatch: "exact"`, an `availableModels` entry allows only the model version it names, so a new release stays blocked until an admin adds it; `deniedModels` blocks specific models outright. The same release adds `/doctor prompt-audit`, which checks CLAUDE.md files, skills, agents, and commands for prompting patterns written for older models.

A developer portal at claude.ai/directory/manage takes plugin bundles from GitHub repositories and remote MCP servers by URL, with no partner program to join first. Every plugin version gets automated validation and a security scan, and a person reviews a new listing before it goes live. After launch, merging to the tracked branch ships an update, and a Usage tab shows installs by surface and version, skill and server runs, and error rates. The older Claude Console submission form is retired.

In a shared Slack channel, Claude Tag can now reach tools connected to one person's account, such as their Google Drive or CRM, alongside the channel's shared connectors. Actions taken through a personal connector appear in that tool's own log under the user's account, and users can review a reply before it posts. Enterprise admins can require review of all personal-connector use, and organizations choose between shared connectors, personal only, or tool by tool. Rolling out on Team plans, with Enterprise to follow.

The Compliance API's local session endpoints are generally available for Claude for Microsoft 365 sessions in Excel, PowerPoint, Word, and Outlook. Separately, the Activity Feed no longer returns file names, project document names, or artifact titles, including on older activities; security tools that relied on those fields need a Compliance Access Key with the `read:compliance_user_data` scope to look them up by ID.

Claude API requests refused before any output are billed again when `stop_details.category` is `bio`, `frontier_llm`, or `reasoning_extraction`, charged at the model's normal rates. Anthropic chose these categories for their low false-positive volume. Early refusals in other categories remain unbilled, and mid-stream refusals were already billed.

OpenAI ended developer access to the Sora video generation API, completing a two-stage shutdown that closed the Sora web and mobile apps on April 26. Integrations that still call the API stop working, and users were told to download their generated content before the cutoff.

Cache diagnostics no longer needs a beta header on the Claude API: include a `diagnostics` object on a Messages request to get a report on why a prompt missed the cache. Responses from `POST /v1/messages` now always carry a `diagnostics` field, set to null when the request did not ask for one. Requests that still send the old beta header keep working.

A single catalog of over 2,000 connectors and plugins, Claude-powered products from companies such as Harvey, Cursor, CrowdStrike, and Snowflake, and consulting firms from the Claude Partner Network. Customers can use a portion of their committed Anthropic spend to buy listed agents, products, and services, which can shorten procurement for organizations with an Anthropic contract.

Meta launched third-generation Ray-Ban Meta glasses at $449 and announced Ray-Ban Meta Audio, its first camera-free glasses, shipping October 13 at $349. It also showed the Muse Charm, a keychain-sized device for talking to its Muse agent by voice, due this fall with no price yet, and VR glasses for spring 2027.

Gemini 3.8 Flash TTS and 3.8 Flash-Lite TTS are generally available in the Gemini API. Developers can design a custom voice from a description or replicate one from a sample, and the voice library grows to more than 150 prebuilt and custom voices.

The first 5.5-family model, priced at $4/$20 per million tokens with cache reads cut 60% to $0.20. Anthropic says it performs at Fable 5.1's level on most work at 40% lower task cost than Opus 5, and it becomes the default Opus in Claude Code. On the API, thinking can no longer be disabled, forced tool use returns an error, and default effort drops to medium.

Two cheaper models in the GPT-6 family below Astra. Sol costs $2/$10 per million tokens, half its predecessor, and OpenAI says it makes about half as many mistakes as GPT-5.6 Sol. Luna, for fast high-volume work, costs $0.10/$0.50. Both are in the API, Codex, and ChatGPT, with Luna reaching Free and Go users.

With the inline-tools-2026-09-15 beta header, a mid-conversation system message can carry a full tool definition, and with mcp-client-2026-09-15 a whole MCP server's toolset. Agents can add or change tools partway through a session without editing the tools array, so the prompt cache and preserved thinking stay intact.

From Claude Code v2.1.278, the classifier that checks risky auto mode actions such as shell commands and network requests runs on Anthropic's servers as part of the session's own model requests, with no charge for the checks. On by default for Enterprise plans and Claude API accounts, and on Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry as each platform rolls out. Gateways that strip unknown request fields force a fallback to the old billed checks; `/status` shows which a session is using, and `CLAUDE_CODE_AUTO_MODE_SERVER=0` opts out.

Claude Code v2.1.277 loads a repository's AGENTS.md — the instructions file other coding agents read — when the project has no CLAUDE.md or CLAUDE.local.md, so a repo set up for other tools works without an import line. A CLAUDE.md in the path still takes precedence by default; a Project instructions setting in /config can load both. Not yet available on Bedrock, Google Cloud, or Foundry, where importing @AGENTS.md from a CLAUDE.md still works.

The Compliance API's local session endpoints now return transcripts of Claude in Chrome sessions, identified by the product_surface value claude_in_chrome. In beta for Claude Enterprise organizations, using the existing Compliance Access Key and read:compliance_user_data scope. Chrome joins Cowork, Claude Code, Claude Science, and Claude for Microsoft 365 as surfaces compliance teams can export for eDiscovery and DLP.

Anthropic and Accenture, through its Faculty division, set up independent evaluation where evaluators work inside the AI company with access comparable to an employee's — red-teaming, safeguard testing, and alignment assessments, and checking whether stated commitments are kept. Each company commits at least $1 billion over five years; Anthropic funds the work and keeps responsibility for model safety. The arrangement is non-exclusive, and Anthropic says more evaluation partners will follow.

Verified life sciences organizations on Team and Enterprise plans get Mythos 5.1, Opus 5, and Sonnet 5 with biology safeguards relaxed across the team (Standard grant, renewed yearly) or removed for a single approved project (High-Risk grant, renewed every six months; Opus 5 and Sonnet 5 only for most applicants). Blocking per request is replaced by monitoring usage patterns, which requires 30-day retention. Available in claude.ai, Claude Science, Claude Code, and the API; not on third-party clouds or BAA-enabled organizations. In beta.

A Claude Project stops being a folder with instructions attached and becomes a standing piece of work. You set the goal and context; Claude proposes tasks, delegates them to threads that run in parallel, and assembles the results. Each thread is a Claude Code cloud session on its own branch — one thread per repo to migrate callers, run tests, and open PRs. Adds project memory that carries decisions forward across conversations, and a Library collecting files and generated artifacts. Beta for select Pro and Max subscribers using cloud sessions, with a waitlist; local execution is not available yet.

A version of GPT-6 Astra with a legal search index covering U.S. case law, statutes, regulations, and court rules across more than 230 million URLs, plus instructions for legal analysis and writing and 26 plugins for tools such as Relativity and Clio. OpenAI reports 40% more accurate research answers than GPT-6 Astra with web search alone. Initially offered to selected law firms through Trusted Access in ChatGPT and Codex, with API access to follow.

OpenAI pilots a new ChatGPT ad format: tapping a sponsored result opens a separate, labeled conversation with an agent run by the advertiser, which answers follow-up questions and hands the user to the merchant's site. Wayfair and Angi are among the launch advertisers. The same update adds natural-language campaign creation and AI creative tools for advertisers, with HubSpot and Shopify as the first CRM and commerce integrations.

Anthropic merges Cowork into Claude. There is no separate surface to open: what Cowork and Claude Design could do is available from any conversation, with existing chats, projects, artifacts, connectors, and skills in place and no migration step. Three betas ship alongside it on paid plans — Claude Docs for writing a document together, Claude Slides for decks you can edit, present from Claude, or download as PowerPoint or PDF, and Claude Design inside conversations. Rolling out to Pro and Max on web, desktop, and mobile, then Free and Team; Enterprise gets 30+ days notice and admins choose when to enable the betas.

Google makes two audio-to-audio models generally available in the Gemini API: gemini-3.8-live, the default for low-latency voice agents, and gemini-3.8-live-extended-thinking, which keeps reasoning in the background while the conversation continues. Aimed at voice agents that have to answer quickly and still work through harder requests.

Gemini in Gmail, Docs, Sheets, Drive, and Chat can now read data from Asana, Atlassian Rovo, HubSpot, Intuit Mailchimp, Intuit QuickBooks, Monday, and Salesforce through MCP integrations. On by default for eligible Business, Enterprise, Education, and Google AI consumer plans, and managed from the Admin console.

Claude for Small Business adds workflows for back-office work (weekly briefs, month-end close, payroll prep) and growth (lead response, proposals, campaigns), with new integrations including Shopify, Square, Stripe, Xero, Gusto, and Zapier. Work is staged for the owner's approval by default. Anthropic also announced free workshops in 10 US cities, 150+ approved trainers, and a learning series with Goldman Sachs 10,000 Small Businesses. Available on all paid plans; 900,000+ installs since the May launch.

Meta launches Meta One, a subscription bundling Facebook Plus, Instagram Plus, and WhatsApp Plus at $7.99/month for Core and $19.99/month for Premium. The headline benefit of both tiers is expanded AI usage, with Premium allowing more than Core. It is Meta's first attempt to charge consumers directly for AI capacity across its apps, alongside separate plans aimed at businesses and creators.

A beta plugin with 37 sales skills that reads Salesforce and Slack to prep calls, score deals, build pipeline dashboards, and draft forecasts. It can update opportunities, log calls, and create tasks in Salesforce, with the seller's approval required by default. Admins connect it once through AgentExchange and pick which groups get access; each seller sees only what their Salesforce permissions allow. Open to paid plans at organizations approved through Salesforce's beta program.

With the compact-2026-09-04 beta header, developers send a top-level compaction parameter and get back one signed summary block, which replaces the summarized messages on later requests. You choose when to compact, the summary can be written in the background while the conversation continues, and recent turns can be kept word for word — with their thinking still valid on models with preserved thinking. Claude API only, on Sonnet 4.6 and newer models.

Connectors for Addepar, Charles Schwab, Orion, SS&C Black Diamond, Envestnet, Vanguard, BlackRock, iCapital, Wealthbox, Zocks, and Wealth.com, plus eight advisor skills: meeting prep, post-meeting notes, rebalance reviews, estate and tax briefs, alternatives briefs, prospect intake, onboarding, and a compliance workflow with Marketing Rule screening. Outputs are drafts for advisor review. Anthropic recommends Enterprise for RIAs because audit logs support recordkeeping.

Claude Code v2.1.269 adds `claude plugin eval`, which runs a plugin's eval suite against Claude Code and returns scored, reproducible results as JSON and an HTML report. It closes the gap that made plugin maintenance guesswork: skills that stop triggering after a description edit, or start triggering on the wrong requests, produce no error and are invisible to code review. The same release ships `/output-style` for listing and switching output styles in Remote Control, cloud, and other headless sessions, and Bash diff output showing which files a command changed when `bashEditDiffEnabled` is set.

DeepSeek releases V4.1 Flash, a multimodal mixture-of-experts model with a 552B-parameter backbone and up to a one-million-token context window, published with MIT-licensed open weights on Hugging Face and served through the DeepSeek API as `deepseek-flash`. Flash API prices drop roughly 11–57% depending on token type, and the previous V4 Flash and V4 Flash Vision Exp models are retired, with their names temporarily routed to the new one. The significance is the price floor: an openly licensed model at this context length changes the build-versus-buy math for teams weighing self-hosting against a frontier API.

Enterprise admins can generate a report from a sample of a team's transcripts over up to 28 days, covering workstreams, cost per session by output type, task outcomes, frictions, the most expensive sessions, and repeated work worth packaging as a shared skill. Individual conversations appear only in aggregate unless an admin turns on attribution, which is logged. The beta includes 10 free reports per organization per month.

A version of ChatGPT Work for banks and investment firms, running on GPT-6 Astra with built-in data from Daloopa, PitchBook, LSEG News, and Crunchbase — no separate data contracts or connectors to set up. It targets research, financial models, and pitch materials, with citations that trace figures to sources, plus role-based access and audit-log exports. Designed with Morgan Stanley and Evercore; available to eligible institutions on request.

GPT-Live-1 brings ChatGPT's full-duplex voice to the API: it listens and speaks at the same time, handles interruptions, and hands harder work off to a backend model. It ships with 12 voices, native transcripts, and WebRTC, WebSocket, and SIP telephony connections. Voice sessions cost $0.05 per minute billed per second, with backend model usage billed separately. OpenAI reports 83.6% task completion on customer-support scenarios when paired with GPT-6 Astra, against 45.7% for GPT-Realtime-2.1.

Anthropic publishes findings from its Claude SMB Tour, which ran hands-on sessions with more than 1,000 small business owners across ten US cities and collected 635 exit surveys. Roughly 80% of attendees ran 5–50 person companies in traditional industries — construction, manufacturing, trades — and many built working custom tools with no software background. Data security was the top pre-adoption barrier, ahead of cost or capability. Nearly two-thirds named guided build sessions the most valuable part, and attendees prioritised hearing from other owners over product demos. Useful as evidence for anyone arguing that AI rollout is a training problem before it is a tooling problem.

OpenAI's Agents API runs agents in the cloud with orchestration, long-running sessions, and context management handled by OpenAI. Agents can run code, work with files, and produce artifacts inside a sandbox that is either OpenAI-hosted or self-hosted on providers such as E2B, Modal, DigitalOcean, and Vercel. There is no platform fee: you pay for tokens, tools, and container time. It is OpenAI's direct counterpart to Claude Managed Agents.

Google releases a desktop Gemini app for Windows 10 and 11. Alt + Space opens Gemini over whatever you are working on, and the app connects to Gmail and Drive and runs multi-step tasks through Gemini Spark. Spark and Gemini Omni media generation require a Google AI subscription. Available globally from gemini.google/desktop.

Managed Agents permission policies gain a third mode. Under `auto`, the server evaluates every agent and MCP tool call against the tool, its input, and the session so far, then runs it, denies it as high-risk, or pauses for approval — and a denial cannot be overridden by the client. Each `agent.tool_use` and `agent.mcp_tool_use` event now carries `evaluated_permission` plus an `evaluation` object naming the policy and a `reason_code`. Anthropic is explicit that `auto` is not a human checkpoint: a call judged safe runs before anyone sees it, so tools that need review still need `always_ask`. The server reads `user.message` content as your intent, which matters if you relay untrusted end-user text. Shipped the same day: `ant beta:sessions connect`, which attaches a terminal to a live session to follow it, interrupt it, and answer waiting approvals.

Anthropic publishes its own cost-reduction guidance for the Claude Platform, with case studies showing 52–73% cost reductions while accuracy held flat or improved. Much of the saving comes from removing text: verification rituals, emphasis boosters, and hand-rolled scratchpads written for 2024-era models that current models no longer need and are slowed by. The other levers are cache hit rate — ordering stable content before variable content so the prefix actually matches — and effort calibration, including testing a stronger model at lower effort against a weaker model working harder. Three Claude Code commands ship with it: `/claude-api prompt-audit`, `/claude-api hillclimb`, and `/claude-api cost-optimize`.

OpenAI ships Images 2.5 across ChatGPT, ChatGPT Work, and Codex, with two API variants: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Flare is the faster default for everyday and high-volume generation, reported at higher quality than GPT-Image-2 with up to 50% lower latency than Images 2.0. Sunburst trades speed for finer control on detailed creative and editing work. The improvements OpenAI highlights are sharper detail, more natural lighting and texture, better preservation of people and products from reference photos, and more reliable editing across a long back-and-forth.

Google starts switching Android phones, tablets, Wear OS watches, paired headphones, and Android Auto vehicles from Google Assistant to Gemini. The rollout runs over several weeks and there is no way to revert to Assistant once a device has been moved. Devices on Android 9 or below with under 2GB of RAM keep a legacy Assistant for basic commands. Assistant has already been removed from ChromeOS.

OpenAI releases GPT-6 Astra, model ID `gpt-6-astra`, with a 1M-token context window at $10 per million input tokens and $50 per million output tokens. OpenAI positions it as state of the art on computer use, browsing, software engineering, cybersecurity, and science, reporting 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench. It is the first OpenAI model to reach the company's Critical internal cybersecurity threshold, so access rolled out in phases — companies in OpenAI's Daybreak cybersecurity program first, then ChatGPT Plus, Pro, Business, and Enterprise, the OpenAI API, Microsoft Azure, and AWS Bedrock. The public version restricts some cybersecurity prompts.

Version 1.30.0 of the ant CLI adds `ant apply`, which creates and updates agents, environments, skills, memory stores, and scheduled deployments from Markdown, YAML, or JSON files in a repository. It prints a plan, waits for approval, and writes a claude-lock.json mapping each file to the resource it created, so later runs update those resources instead of duplicating them. Resources reference each other by relative path and are applied in dependency order. If a resource was changed in the Console since the last run, the plan refuses to apply rather than overwriting it. `--dry-run` posts the plan on a pull request; `--yes` applies after merge.

Meta releases Muse Spark 1.3, its frontier reasoning model, through the Meta Model API and the Muse Code coding agent. It has a 1M-token context window and takes text, image, and video input. The standard endpoint, where prompts stay private, runs $1.25 per million input tokens and $4.25 per million output tokens. A separate contributor endpoint costs roughly $0.10 / $0.20 per million — around 10-20x cheaper — in exchange for letting Meta train on your traffic. That split is the notable part: it makes training consent an explicit line item on the price sheet.

Anthropic publishes a commerce agent blueprint with two runnable reference implementations. A shopping agent searches a catalog, assembles multi-item carts, completes checkout, and answers order and policy questions in conversation. A merchant agent handles sales analytics, inventory tracking, pricing recommendations, and drafted campaigns, with a person approving before anything goes live. Deployable on the Claude API, Bedrock, Microsoft Foundry, and Vertex AI, via the Messages API, Agent SDK, or Managed Agents. Anthropic reports enterprise customers seeing carts up to 35% larger and shoppers 60% more likely to complete a purchase.

Google's third Flash release in six weeks, after 3.6 Flash in July and 3.7 Flash on August 13. Gemini 3.8 Flash improves on 3.7 Flash in software engineering, agentic tasks, and multi-step reasoning, at an introductory $0.75 / $3.75 per million input / output tokens through December 31, 2026. It is available in Google AI Studio, Android Studio, Gemini Enterprise, and the Gemini apps. The 3.8 Flash Cyber variant, tuned for vulnerability detection and automated patching, is limited to government and critical-infrastructure defenders through Google's Fairwind Program.

Fable 5.1 (claude-fable-5-1) ships at the same $10/$50 per million tokens as Fable 5, with cache reads cut from $1.00 to $0.25 per million — 0.025× base input, against 0.1× on every other Claude model. Typical workloads land about 25% cheaper and heavily agentic ones up to 45% cheaper. It scores 52.6% on Terminal-Bench-Science against Fable 5's 24.7%, and 31.4% on AutomationBench against Opus 5's 26.9%. Three changes break existing code: forced tool use now returns a 400, thinking blocks are bound to the model that produced them, and editing earlier turns invalidates them. Mythos 5.1 is the same model with safeguards loosened for Project Glasswing participants.

Anthropic announces Enterprise Frontier Safeguards, which resolves the conflict between zero data retention and cross-session misuse detection that had locked regulated customers out of Fable 5. Activity data used for monitoring lives in the customer’s own AWS, Google Cloud, or Azure account under their own encryption keys and never reaches Anthropic’s servers. Anthropic’s automated systems analyse it for attempts to build offensive cyber or biological capability and for signs of stolen credentials; detections route to the customer’s own security team, with no Anthropic personnel reviewing content. Designed with more than 100 enterprise customers, including roughly a quarter of the Fortune 100. Rolling out in phases, request-based, with broad availability targeted for later in autumn 2026.

Claude for Teachers expands from individual educator accounts to a free Enterprise offering for entire US K-12 schools and districts. School and district leaders verify their institution and manage access centrally for educators and staff; teachers who had verified individually transition into the district account automatically. Two new teaching skills ship — lesson preparation and check for understanding, maths-focused at launch — alongside standards alignment across all 50 states. Districts signing up by June 30, 2027 get a full year free, with overage billing off by default.

The Claude in Chrome extension leaves pilot and becomes generally available on every paid Claude plan. Claude reads the page you are on and acts on it — typing, clicking, navigating, filling forms — using your existing logins. The change at GA is that it can now automatically approve actions a safety classifier judges safe, rather than prompting on every one; manual approval remains available as an opt-in. Three defence layers cover prompt injection: model training against a red-team attack library, content probes that scan pages before Claude acts, and a classifier that blocks actions not matching the original request. Chrome only, and enterprise admins can restrict it to approved domains.

Claude Cowork gains a dedicated browser inside the desktop app that Claude drives on its own. It never touches your tabs, bookmarks, or passwords — you choose which individual logins to transfer across. Built for delegated web work: pulling invoices from vendor portals, gathering research, filling forms while you do something else. Available on Pro, Max, and Team, and immediately on Enterprise with admin control through Organization settings. macOS, Windows, and Linux, with Windows and Linux in beta. Banking, email, and single sign-on sites are restricted by default.

Salesforce and Anthropic announce Claudeforce, an expanded partnership making Claude the default reasoning model across Agentforce (the Atlas Reasoning Engine, Agentforce Vibes, Agentforce Coworker, and Agent Builder), Slack AI and Slackbot, and Headless 360. It launches with Salesforce in Claude, a plugin carrying 37 prebuilt sales skills for meeting prep, deal health review, and pipeline review, plus an onboarding dashboard over live CRM data. Available to select pilot customers, with open beta expected in September 2026. Also offered inside the Salesforce Trust Boundary via Amazon Bedrock for regulated industries.

Google releases Gemini 3.5 Transcribe in public preview, a speech-to-text family built on Gemini’s audio understanding and positioned as the successor to Chirp 3. Two variants: gemini-3.5-transcribe for pre-recorded files up to an hour with speaker diarization and word-level timestamps, and gemini-3.5-transcribe-live for real-time streaming over WebSockets in 10-minute sessions. Google cites Artificial Analysis word error rates of 2.6% non-streaming and 4.0% streaming. Priced at $0.005 per minute for files and $0.009 for live, undercutting Google’s own Cloud Speech-to-Text v2 standard tier.

Memory stops being per-surface. The memory built in chat is the same one Cowork uses, including Cowork tasks running in the cloud, and it works across web, desktop, and mobile. Claude now captures information as conversations happen rather than summarising afterwards. Users view, edit, and delete individual memory topics in Settings, and can pause or reset memory entirely. Sensitive subjects — health, beliefs, identity — are opt-in only, and identification numbers, criminal history, and immigration status are never stored. On by default for Free, Pro, and Max; Team and Enterprise users enable it themselves once an admin allows it.

Anthropic widens access to Claude Mythos 5, the model whose safety classifiers are loosened for offensive-security research, extending it to more vetted defensive cybersecurity organisations through Project Glasswing. The model is the same as Fable 5 underneath; what differs is the classifier configuration, which allows work — vulnerability analysis, exploit documentation, threat modelling — that the general-availability model declines.

Anthropic launches Claude Academy at academy.claude.com, consolidating its training material into one platform: 22 courses, 119 tutorials, 148 role-specific use cases, and live webinars, all free. anthropic.com/learn now redirects there. The platform adds badges, completion tracking, personalized recommendations, and a Claude Academy Skill that suggests courses. Anthropic’s stated philosophy is mindsets over features — durable habits like verifying in proportion to the stakes, rather than product behaviours that go stale.

Four beta headers drop at once on the Claude API. Computer use becomes computer_toolset_20260801 with batch actions, zoom on by default, and per-member configuration. A new browser use toolset drives a browser your application hosts, reading the page’s accessibility tree and offering element references, form input, and tab management rather than pixel-coordinate clicking. Agent Skills and the Skills API go GA, as does the Files API with file expiration and pagination. Supported on Fable 5, Mythos 5, Opus 5, Sonnet 5, and Opus 4.8.

Workbench in the Claude Console is replaced by Playground, which supports every Messages API parameter, ships templates demonstrating code execution and web search, and shows the full SDK request alongside the API response for each run. The legacy Workbench and prompt tools APIs retired the previous day, so any automation calling them has already broken.

New Claude Code sessions on Pro, Max, and Team plans now start in auto mode, where an AI classifier approves commands it judges safe instead of prompting for each one. Enterprise, the Claude API, Bedrock, Google Cloud, and Microsoft Foundry stay opt-in. Admins can set defaultMode or remove the option entirely with disableAutoMode. Anthropic notes the classifier does not eliminate risk and recommends manual review for production infrastructure changes.

Claude Tag — Claude working as a tagged participant in Slack — gains fuller channel context, so a question asked mid-thread is answered against what the channel has already discussed rather than the tagged message alone. Anthropic published its own internal deployment alongside it, using Tag for self-service data analytics on ad-hoc questions.

Google ships gemini-3.7-flash to general availability, positioning it as its most intelligent workhorse model for coding and agents, with improvements across software engineering, web development, and agentic workflows. Introductory pricing runs through December 31, 2026. It follows Gemini 3.6 Flash and 3.5 Flash-Lite, which went GA on July 21.

Anthropic folds the Claude in Chrome side panel into Cowork, so the browser surface now carries Cowork’s task delegation and saved conversations rather than being a separate lightweight assistant. One fewer product surface to explain to a rollout, and one more place Cowork usage shows up in admin analytics.

The Compliance API can now retrieve transcripts of Cowork and Claude Code sessions that run locally on employees’ machines, in beta for Claude Enterprise. Three endpoints under /v1/compliance/apps/sessions/local list sessions, fetch metadata, and return transcripts, using the existing Compliance Access Key. It closes the last major blind spot for organizations that need a complete record of AI activity. A companion change adds an anthropic-workspace-id response header to the API.

The $2 / $10 per million token introductory price for Claude Sonnet 5 is now the standard price. The increase to $3 / $15 scheduled for September 1, 2026 will not happen. Sonnet 5 is the model most production agent fleets run on, so a cost model built against the September increase has just gained back a third of its headroom.

Four controls land on Claude Managed Agents. A session budget sets a hard cap on spend — a session that reaches it pauses with a budget_reached stop reason instead of starting new model requests. An advisor entry in the multiagent roster lets the primary thread consult a stronger model mid-turn for strategic guidance. inference_geo pins where inference runs, per agent or per session. And sessions can load Skills straight from a mounted GitHub repository’s .claude/skills directory. On August 19, self-hosted sandboxes also gained memory stores, and the Console session viewer was rebuilt with per-tool statistics.

Claude Enterprise gains automatic security scanning of third-party skills and plugins: when someone uploads or edits one, it is checked for malicious content before it can run. Skills and plugins are packaged instructions and tools that extend what Claude can do, and until now nothing inspected them on the way in. In beta on Enterprise plans.

Claude Code sessions can run in an environment your organization hosts rather than Anthropic’s, in public beta. It unblocks teams that could not adopt cloud-run sessions for data-residency reasons, and pairs with the self-hosted sandboxes shipped for Managed Agents in June.

Claude Enterprise organizations can route every governed prompt through a security server they run, which returns an allow or deny verdict before inference begins. A denied request never reaches the model. The hook runs on Anthropic's servers and covers claude.ai, Cowork, and Claude Code, with shadow mode, a rollout percentage, and role exclusions for staged deployment. In beta; not available on Bedrock or Google Cloud.

Anthropic retires Claude Opus 4.1 (`claude-opus-4-1-20250805`). Requests to the model now return an error rather than falling back. Anthropic recommends Claude Opus 5, which sits at the same $5/$25 per million tokens as Opus 4.8. Researchers can request continued access through the External Researcher Access Program.

OpenAI publishes "Ten advances in mathematics and theoretical computer science," reporting that an internal version of Astra — its next major model — produced new constructions, counterexamples, and bounds for ten problems that had seen no progress for at least a decade. Each argument is formalized as a machine-checkable Lean proof, published in the openai/ten-proofs repository, so verification does not depend on trusting the model. OpenAI put the token cost of finding all ten at roughly $2,000 at Sol API rates.

Anthropic ships support for the 2026-07-28 revision of the Model Context Protocol across Claude. MCP is the open standard — now governed by the Agentic AI Foundation — that lets Claude connect to external tools and data sources through a common interface rather than a bespoke integration per system. Spec revisions matter to anyone maintaining an MCP server: each one sets the contract that servers and clients agree on, and lagging behind a revision is how connectors quietly stop working.

Anthropic and Cognizant expand their partnership to bring Claude to Cognizant's enterprise client base. It follows the same pattern as the KPMG alliance (May 19) and the Blackstone/Goldman Sachs deployment joint venture (May 4): Anthropic reaching enterprises through systems integrators rather than direct sales alone. For IT leaders evaluating Claude, it increases the odds your existing consulting partner already has a Claude practice.

Claude Opus 5 (`claude-opus-5`) launches as a step-change over Opus 4.8 at unchanged pricing of $5/$25 per million tokens — half the cost of Fable 5, with Anthropic positioning it as close to Fable 5's frontier intelligence. It ships with a 1M token context window as both default and maximum, 128k max output, thinking on by default, and the full effort ladder (`low` through `max`), with effort as the primary control for reasoning depth. The prompt cache minimum drops to 512 tokens. One breaking change: `thinking: {"type": "disabled"}` at `xhigh` or `max` effort now returns a 400. Two behaviours to watch when migrating — `max_tokens` now has to cover thinking tokens, and the model self-verifies unprompted, so verification instructions carried over from older models cause over-verification. Fast mode for Opus 4.7 was removed the same day and now errors rather than falling back.

Claude voice mode moves onto the Opus and Sonnet model families, adds tool actions against connected Gmail and Slack accounts, supports eleven languages, and lets you switch models mid-conversation. Voice moves from a transcription front-end to something that can act on connected systems while you talk.

Anthropic launches four role-based certifications for people who deploy Claude for customers, alongside a voice mode update that lets users think through hard problems by talking rather than typing. The certifications are the first formal credential Anthropic has offered around Claude deployment — relevant to the forward deployed engineer and agent manager roles that consultancies and enterprises are now hiring for, where 'how do I prove I can do this' has had no standard answer.

The legacy Workbench at platform.claude.com/workbench is being sunset, with access ending August 17, 2026. Saved prompts, variables, and evals are not carried over to the replacement Workbench at /playground — you have to export anything you want to keep from the banner or Organizational Settings. The experimental prompt tools APIs retire on the same date: `/v1/experimental/generate_prompt`, `/improve_prompt`, and `/templatize_prompt` will return errors. Anything in your tooling that calls those endpoints needs replacing before mid-August.

Claude Enterprise organisations can now manage people programmatically through the Admin API, in beta: list members, look them up by email, change roles, remove members, send and withdraw invites, manage groups and membership, and read custom roles. Group and custom-role calls need the `ce-user-management-2026-07-13` beta header; member and invite calls need none. This is the piece that lets identity automation — joiner/mover/leaver flows, offboarding scripts — cover Claude the same way it covers other SaaS. Anthropic also launched Claude for Teachers the same day.

OpenAI opens GPT-5.6 to the public two weeks after its government-gated preview, releasing all three tiers — Sol (flagship), Terra (mid-range), and Luna (fast and cheap). Alongside it comes ChatGPT Work, an agent built to carry out whole jobs rather than answer questions, and a small-business programme bundling ChatGPT Work and GPT-5.6 with training and in-person AI academies. OpenAI also shipped GPT-Live, a full-duplex voice generation that listens and speaks simultaneously. The public release lands two weeks before Claude Opus 5.

Meta ships Muse Image inside Meta AI, its first image generation model from Meta Superintelligence Labs. It uses reasoning to interpret complex prompts and blends multiple source photos into a single output. Rolling out to Facebook, Messenger, Instagram, and WhatsApp. It follows Muse Spark (April) as the second proprietary model from the lab, continuing Meta's move to run closed models alongside the open Llama line.

Claude Enterprise admins get usage and cost broken down by group and by individual user, an Analytics API that feeds tools like Datadog Cloud Cost Management and CloudZero, model defaults and entitlements that set which model conversations start on across chat, Cowork, and Claude Code, and spend threshold alerts firing at 75% and 90% of an org limit. Users get their own warnings at 75% and 95% plus an in-app way to request an increase before they hit the cap.

Meta is building Meta Compute, a cloud infrastructure unit that will sell spare AI computing capacity to outside customers, putting it in direct competition with AWS, Microsoft Azure, and Google Cloud. It is the clearest sign yet that the AI buildout has produced more capacity than any single company's own products consume, and that compute itself is becoming the product. For enterprises, another potential venue for running models — and another vendor in an already crowded procurement conversation.

The US Commerce Department lifted the export controls on June 30, and Anthropic put Claude Fable 5 and Claude Mythos 5 back online worldwide on July 1 — across Claude.ai, the Claude Platform, Claude Code, and Cowork. To get there Anthropic shipped a new safety classifier that blocks the underlying jailbreak at a 99%+ rate, agreed to four government-collaboration terms, and — with Amazon, Microsoft, Google, and other Project Glasswing partners — proposed a shared jailbreak severity framework that scores each jailbreak on factors like capability gain (how far it takes a user beyond tools they already have). The three-week round trip is the case study for model-continuity planning: a frontier model can vanish and return on government action, so production systems need a wired-in fallback model and refusal handling regardless of which lab they build on.

Anthropic launches Claude Science, a beta workbench (macOS and Linux; Pro, Max, Team, Enterprise) that pulls a researcher's scattered tools into one place — 60+ scientific databases and toolkits spanning genomics, single-cell work, proteomics, structural biology, and cheminformatics. Ask a question in plain language and a coordinating agent hands work to specialist sub-agents while a separate reviewer agent checks every citation and calculation as it goes. The bet is on workflow, not a new model: it runs on existing Claude models but packages the multi-agent orchestration and auditable outputs that regulated research demands. Anthropic is funding up to 50 'AI for Science' projects with up to $30k in credits, with applications open through July 15.

Google opens its Gemini 3.5 family with Gemini 3.5 Flash — a faster, lower-cost model for coding, agents, multimodal tasks, and long-document work — now with built-in Computer Use that lets agents see, reason, and act across desktop, mobile, and browser environments. Alongside it, Gemini Omni Flash (a natively multimodal model for building custom video workflows) reaches public preview and Nano Banana 2 Lite ships as Google's fastest, most cost-efficient image model. Google signaled Gemini 3.5 Pro is still on the way. The Computer Use integration puts Google head-to-head with Anthropic's and OpenAI's agentic desktop-control offerings.

Claude Sonnet 5 (`claude-sonnet-5`) launches as a drop-in upgrade to Sonnet 4.6 and the new default on Free and Pro, with its biggest gains in coding and agentic work — on Terminal-Bench 2.1 it even edges the Opus 4.8 flagship (80.4% vs 74.6%). It ships with a 1M token context by default, 128k max output, and adaptive thinking on by default, and is the first Sonnet-tier model with real-time cybersecurity safeguards (refused requests return a 200 with `stop_reason: 'refusal'`). Introductory pricing is $2/$10 per million tokens through Aug 31, 2026 (standard $3/$15). Three migration gotchas beyond the model-ID swap: manual extended thinking (`budget_tokens`) now returns 400, non-default sampling params (`temperature`/`top_p`/`top_k`) now return 400, and a new tokenizer produces ~30% more tokens for the same text — so recount prompts and revisit `max_tokens` and cost budgets before switching.

OpenAI previews the GPT-5.6 family — Sol (flagship), Terra (balanced, ~2x cheaper than GPT-5.5), and Luna (fast, lowest-cost) — priced at $5/$30, $2.50/$15, and $1/$6 per million tokens. At the US government's request, access starts as a limited preview for roughly 20 trusted partners (whose participation was shared with the Administration) through the API and Codex, not ChatGPT, with broader release planned 'in the coming weeks.' It closely mirrors Anthropic's Fable 5 episode: frontier releases are now routed through government pre-release review. OpenAI also plans to run Sol on Cerebras hardware at up to 750 tokens/sec, and retired GPT-4.5 from ChatGPT the same day.

Anthropic launches Claude Tag (beta, Team and Enterprise), moving the unit of AI collaboration from the private conversation to the shared Slack channel. One Claude joins the workspace as a persistent teammate with its own identity, its own memory, and admin-scoped access to tools and data. Tag it on a task and it breaks the work into stages and executes in the background, posting results back to the channel; with 'ambient' mode enabled it proactively flags relevant information and follows up on threads that have gone quiet. It replaces the old Claude-in-Slack integration (retiring Aug 3), with a 30-day window to migrate. A concrete template for the 'agent as coworker' pattern — Anthropic says its internal version now writes about 65% of its product team's code.

Claude Code gains artifact support, letting a session publish a rendered page — a report, a dashboard, a diagram — rather than leaving output as terminal text or a file on disk. It closes a gap between Claude Code and the web app for work whose deliverable is something a person needs to look at and share rather than run.

The original Claude 4 models — `claude-sonnet-4-20250514` and `claude-opus-4-20250514` — are retired, and all requests to them now return an error. Anthropic recommends Sonnet 4.6 and Opus 4.8 as replacements. Researchers can request continued access through the External Researcher Access Program. The retirement was announced with the Opus 4.7 launch in April, giving roughly two months of notice — a useful data point for how much runway to expect on future deprecations.

Three days after launch, Anthropic receives a US government export-control directive (June 12, 5:21pm ET) barring access to Fable 5 and Mythos 5 by any foreign national — inside or outside the US, including foreign-national Anthropic employees. To comply, Anthropic abruptly disables both models for all customers worldwide. The stated concern is national security tied to a method of bypassing Fable's safeguards; the directive gives no specifics, and Anthropic — which reviewed a demo tied to a few previously known minor vulnerabilities — calls it a misunderstanding and says it is working to restore access. All other Claude models (Opus 4.8, Sonnet, Haiku) are unaffected. A live lesson in cloud-AI continuity: a hosted model can vanish overnight, so production systems need a fallback model wired in.

Anthropic launches Claude Fable 5 (`claude-fable-5`), its most capable widely released model, built for demanding reasoning and long-horizon agentic work. It ships with a 1M token context window by default, up to 128k output tokens, and always-on adaptive thinking (the older on/off thinking toggle is gone). Pricing is $10/$50 per million tokens — double Opus 4.8. It posts 95.0% on SWE-bench Verified and 80.3% on SWE-bench Pro (an 11-point jump over Opus 4.8). The integration-breaking change: built-in safety classifiers can decline a request, returning `stop_reason: "refusal"` as a successful 200, so apps need refusal handling and a fallback path to another model. Claude Mythos 5 (`claude-mythos-5`) shares the same capabilities without the classifiers and is limited to Project Glasswing partners. Free on Pro/Max/Team/Enterprise through June 22, then metered against credits.

Anthropic expands Claude Managed Agents along three lines. Self-hosted sandboxes reach general availability on the Claude API and Claude Platform on AWS: tool execution runs on infrastructure you control and can reach your private MCP servers, while Claude keeps running on Anthropic's side. Agents can now run on a schedule without an external trigger, and environment variables (API keys, credentials) can be stored in encrypted vaults instead of being passed in per request. Together these move Managed Agents closer to production use inside enterprise security boundaries.

OpenAI confidentially submits a draft S-1 registration to the SEC — the legal step that precedes an initial public offering. Reported valuations run $730B–$852B against annualized revenue above $25B. The company says it has not set a timeline and that some plans are easier to pursue while still private. It follows Anthropic's own confidential IPO filing in late May, signaling that the two largest frontier labs are both moving toward public-market scale and scrutiny.

Anthropic notifies developers that Claude Opus 4.1 will be retired on the Claude API. With Opus 4.8 and now Fable 5 available, teams still calling `claude-opus-4-1` need to migrate to a current model. A reminder that pinning a specific model ID carries an end-of-life clock — production code should plan for model migration as routine maintenance.

OpenAI ships a new ChatGPT memory system it calls Dreaming, which reviews past conversations to keep preferences and context fresh across sessions. The same week it expands Codex with role-specific updates aimed at non-developers — analysts, marketers, operators, and designers — who now make up a fast-growing share of Codex's 5M+ weekly users. Both moves push ChatGPT further from a chat box toward a persistent, cross-role work assistant.

Mark Zuckerberg unveils Meta Business Agent, an AI agent businesses can deploy across WhatsApp, Messenger, and Instagram to answer customer inquiries, recommend products, and book appointments. It ships inside a business tier of Meta One and is billed on a consumption basis for large WhatsApp Business Platform users. Meta's clearest move yet to monetize AI beyond advertising. Separately, Meta's Muse Spark model API for developers slipped again with no firm launch date.

At Build 2026, Microsoft introduces Autopilots — a new category of always-on, autonomous agents — starting with Scout, which works in the background across Teams, Outlook, OneDrive, SharePoint, browsers, and local files and acts without being prompted each time. Microsoft also ships seven new in-house MAI models (headlined by MAI-Thinking-1, a 35B reasoning model with a 256K context built without distillation) and Copilot Platform, a set of APIs and a runtime for embedding Copilot agents in any app. The framing shift: from AI-as-assistant to AI-as-operator.

OpenAI makes its frontier models and Codex generally available on AWS through Amazon Bedrock, and expands Codex with role-specific plugins, Sites (sharing interactive workspace apps and websites), and annotations for refining documents, spreadsheets, and slides. Codex now has more than 5 million weekly users, and non-developers — analysts, marketers, operators, designers, bankers — make up about 20% of them, growing more than 3x faster than developers.

Shipping alongside Opus 4.8 as a research preview, Dynamic Workflows let Claude author a multi-step plan and orchestrate work across tens to hundreds — up to roughly a thousand — subagents running in the background. Built for jobs a single context window can't hold: codebase-scale migrations across hundreds of thousands of lines, sweeping audits, large repetitive transforms. Distinct from sub-agents, Agent Teams, and the `claude agents` terminal mode — here Claude decides the structure and you approve the plan rather than each step.

Anthropic closes a $65B Series H round at a $965B post-money valuation, with run-rate revenue reported to have crossed $47B earlier in May. The same day, it confidentially files for an initial public offering — setting up a public-market test of the AI boom. Together they mark Anthropic's transition toward public-company scale and scrutiny.

Anthropic launches Claude Opus 4.8 (claude-opus-4-8), its most capable generally available model. Pricing is unchanged at $5/$25 per million tokens. The effort parameter now defaults to high across all surfaces, a 1M token context window ships by default, and max output rises to 128k tokens. Fast mode runs ~2.5x faster at $10/$50 — three times cheaper than prior Opus fast mode. New mid-conversation system messages let instructions change without breaking the prompt cache. Anthropic reports it beats GPT-5.5 on 12+ benchmarks (SWE-Bench Pro 69.2% vs 58.6%) and is ~4x less likely than 4.7 to let a code flaw pass in review. Migration from Opus 4.7 is a model-name swap — no new breaking changes.

Claude Code v2.1.149 ships two practitioner-relevant features: /usage, a new command that shows per-category token cost breakdowns mid-session (context vs. tools vs. output vs. cache); and allowAllClaudeAiMcps, an enterprise setting that blanket-approves all Claude.ai cloud MCP connectors without per-server admin review. Also in the v2.1.147-v2.1.152 range: /simplify renamed to /code-review, /code-review --fix auto-applies findings, Auto mode runs without a per-session consent prompt, and VS Code 1.121 adds remote agent sessions that survive laptop disconnect.

Anthropic publishes the first update on Project Glasswing — its controlled-access cybersecurity initiative using the unreleased Claude Mythos Preview model. In the first month, Glasswing partners discovered 10,000+ software vulnerabilities. The open-source scan alone (1,000+ widely used projects) flagged 6,202 high/critical vulnerabilities with a 90.6% confirmed true-positive rate. Partners include AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks, and now Verizon.

Anthropic ships the Claude Compliance API and 28 launch-partner integrations across DLP (Microsoft Purview, Nightfall, Forcepoint), SIEM (Splunk, Securonix, Sumo Logic), CASB (Netskope, Zscaler), identity (Okta, CrowdStrike Falcon), eDiscovery (Relativity, Logikcull), and AI security posture (Wiz, Orca, Lacework). Enterprise admins get programmatic access to all Claude usage data — request-level prompts and responses — for compliance, eDiscovery, and security monitoring. The Compliance API and Security Ingestion API are distinct: compliance routes data to legal/HR, security routes alerts to SOC tooling.

Elon Musk's XAI (backed by SpaceX compute infrastructure) acquires Cursor at roughly $60 billion with a $10 billion opt-out clause. The compute access enables Composer 2.5, a new coding model that is significantly faster and cheaper than equivalents on Claude or Codex — early reports put it at near-frontier quality for front-end work at a fraction of the token cost. Cursor announces intent to become a full 'coding and knowledge work' platform, adding automations and a marketplace alongside its existing in-app browser.

OpenAI ships two meaningful Codex updates: /goal mode, which lets agents run autonomously for hours (or more than a day) toward a stated high-level objective rather than stopping after each command; and plugin sharing, which lets team members publish workflow plugins to the entire workspace. Design mode (screenshot + annotation for UI iteration) also ships. Together these move Codex significantly closer to a unified super-app that handles both coding and long-horizon knowledge work.

Google I/O 2026's AI announcements: Gemini Spark (leaked/previewed) — a new agentic surface inside the Gemini app with virtual computers, long agent trajectories, canvas, deep research, and connections across Google Workspace. Gemini 3.5 Flash releases but falls short of frontier expectations — widely noted as faster but not competitive with GPT-5.5 or Claude Opus 4.7 on complex reasoning. Gemini Omni video model also announced but trails Sora and Kling. Google's main challenge: no clear unified super-app strategy (Gemini vs. AI Studio vs. Project IDX).

KPMG signs a global alliance with Anthropic to deploy Claude across its 276,000-person workforce and embed Claude Cowork and Managed Agents into KPMG's Digital Gateway platform. Launch use cases: tax, legal, private equity. KPMG also announces KPMG Blaze — a product built on Claude Code for modernizing legacy IT systems inside private equity portfolio companies. Second Big Four firm to go all-in on Claude in one month, following PwC's announcement on May 14.

Karpathy — founding OpenAI member, former Tesla Autopilot lead, and the foremost AI educator on the internet — announces he has joined Anthropic to get back to R&D. The hire is widely read as a significant talent signal: CTOs of publicly traded companies and top researchers across the industry have been choosing individual contributor roles at Anthropic over running organizations. Karpathy's publicly stated focus: the next few years of LLM development will be "especially formative."

Anthropic ships `claude agents`, a new terminal mode that lets users run multiple independent agent tasks simultaneously. Each task runs as its own agent; keyboard navigation moves between them. Distinct from the desktop parallel sessions sidebar — this is the headless/terminal-native equivalent. Enables parallel research, builds, and pipelines without GUI.

Anthropic acquires Stainless for $300M+. Stainless automates generation of SDKs, CLIs, and MCP servers from API specs, and previously generated all official Anthropic, OpenAI, Google, Cloudflare, and Meta SDKs. Bringing Stainless in-house means Anthropic now controls the SDK pipeline and MCP server generation toolchain. All hosted Stainless products are being wound down; existing customers keep their generated SDKs.

Anthropic announces that from June 15, 2026, programmatic Claude usage (the Agent SDK, `claude -p` headless mode, Claude Code GitHub Actions, and third-party apps using Agent SDK credentials) stops drawing from a subscription's rate limits and instead draws from a separate monthly dollar credit metered at API list prices — $20 for Pro, $100 for Max 5x, $200 for Max 20x, per user, no rollover. Interactive Claude Code and web chat are unchanged. The change ends uncapped subsidized automated usage and is effectively a large price increase for heavy programmatic workloads.

OpenAI launches a majority-owned deployment company with McKinsey, Bain, and Capgemini, funded at $10B+. Simultaneously acquires Tomoro, a 150-person forward deployed engineering firm, to staff it immediately. The Deployment Company puts OpenAI engineers inside enterprise clients to build production AI systems against their actual infrastructure — a direct copy of Palantir's FDE model. Launched within days of Anthropic's $1.5B JV with Blackstone and Goldman Sachs.

Anthropic launches Claude for Small Business: a bundle of pre-built connectors and ready-to-run workflows targeting companies without a dedicated IT or AI team. Seven integrations at launch: QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, Microsoft 365. Claude can read live data from connected tools and perform tasks without manual copy-paste. Available on Claude Pro and higher.

Google's Android Show 2026 centers on Gemini Intelligence: on-device AI that automates complex tasks, summarizes web content, fills forms, and includes Rambler (polishes voice messages or builds custom widgets from natural language). New Googlebooks AI-first laptops announced. Gemini comes to Chrome. Rolling out to Samsung and Pixel phones first, then broader Android. Signals the shift from Gemini-as-chatbot to Gemini-as-OS-layer.

Anthropic's native Claude Platform experience arrives on AWS as a GA service — the first cloud provider to offer it. AWS customers get the full Claude API (Messages, Files, Batches, Managed Agents, Agent Skills, code execution, web tools) through native AWS endpoints, unified with existing AWS billing, authentication, and security controls. No separate Anthropic contract needed. Anthropic simultaneously commits $100B+ to AWS over ten years.

Anthropic ships three new capabilities for Managed Agents. Dreaming: agents review their own past sessions, extract patterns, and self-improve over time. Outcomes: a separate grading agent scores completed tasks and re-runs them until they hit quality thresholds — lifted document generation quality 10.1% on benchmarks. Multi-agent orchestration: a lead agent breaks jobs into pieces and delegates each to a specialist with its own model, prompt, and tools running in parallel on a shared filesystem.

Anthropic ships 10 ready-made agent templates aimed at financial services workflows: Pitch builder, Meeting preparer, Earnings reviewer, Model builder, Market researcher, Valuation reviewer, General ledger reconciler, Month-end closer, Statement auditor, and KYC screener. Each is a reference architecture bundling skills (task instructions + domain knowledge), connectors (governed access to data), and subagents. Available as plugins for Claude Cowork and Claude Code, and as cookbook snippets for Claude Managed Agents.

OpenAI ships GPT-5.5 Instant as the default model for all ChatGPT users, replacing GPT-5.3 Instant. Produces 52.5% fewer hallucinated claims on high-stakes prompts (medicine, law, finance). Can search past conversations, files, and Gmail for personalized answers. GPT-5.3 Instant remains available to paid users for three months before retirement.

Anthropic announces a new joint venture with Blackstone, Goldman Sachs, and Hellman & Friedman to build an enterprise AI services company. The venture will help large companies integrate Claude into core business operations — sales, finance, ops, customer support — with funded engagement teams and prebuilt playbooks. Signals Anthropic going further down the services + integration path, not just the model + API path.

Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs announce a new AI services company targeting mid-sized businesses. The joint venture deploys Claude into core operations across sectors — the first major move by Anthropic to own the services layer, not just the model.

Microsoft ships Agent 365 ($15/user/mo) and the Microsoft 365 E7 "Frontier Suite" ($99/user/mo, bundling M365 E5 + Copilot + Agent 365). Agent 365 is a control plane that discovers, governs, and secures AI agents across an enterprise — Microsoft's own Copilot agents, third-party SaaS agents, AWS Bedrock and Google Cloud agents, plus local agents on Windows endpoints including Claude Code, GitHub Copilot CLI, and OpenClaw. Aimed at IT/security teams managing "agent sprawl."

Anthropic moves Claude Security from research preview into public beta for Claude Enterprise. Built on Opus 4.7. Scans repositories, traces data flows across files and modules, and produces vulnerability findings with proposed patches you review and apply. Three scan modes (scheduled, targeted, ad-hoc); findings export to Slack, Jira, CSV, or Markdown. Started life as Claude Code Security in February. Team and Max access on the roadmap.

OpenAI rolls out Advanced Account Security for ChatGPT: phishing-resistant sign-in, tighter recovery controls, shorter sessions, login alerts, and automatic exclusion from training. Required by June 1, 2026 for individuals using OpenAI's most cyber-capable models under its Trusted Access for Cyber program. Tightens the consumer surface that has lagged enterprise IDP integrations on most LLM products.

Microsoft, Google, Meta, Amazon, and others are on track to collectively spend ~$700 billion on AI infrastructure in 2026, nearly double 2025 levels. Meta alone guided $115–135B in capex. No clear ceiling: every major hyperscaler is accelerating, treating AI compute as the defining infrastructure race of the decade.

The claude-api skill — an open-source bundle of current API patterns, model parameters, prompt caching strategies, and SDK updates — now ships inside JetBrains IDEs (with Junie), Warp terminal, CodeRabbit PR review, and Resolve AI. Originally released in Claude Code in March, the skill keeps developer tools current with new model launches like Opus 4.7 without manual doc lookups. Reduces stale-API errors in production code and surfaces guided migrations inline.

Anthropic launches nine connectors for professional creative software: Adobe Creative Cloud, Blender, Ableton, Autodesk Fusion, Affinity by Canva, SketchUp, Resolume, and Splice. Claude can now tutor, extend via scripting, and bridge across creative pipelines. Three university partnerships (RISD, Ringling, Goldsmiths) anchor an education push. Distinct from Claude Design (April 17), which generates visuals on its own.

Microsoft and OpenAI renegotiate their partnership: OpenAI can now license its models across multiple clouds including Google Cloud and Amazon Web Services. Microsoft loses exclusive IP access in exchange for a capped revenue share through 2030 and a nonexclusive license to OpenAI IP through 2032. Removes a structural barrier that had kept GPT models off competing cloud platforms. Amazon confirms it will make OpenAI models available on Bedrock shortly.

OpenAI ships GPT-5.5 and GPT-5.5 Pro: a single unified architecture handling text, images, audio, and video end-to-end. Co-designed with NVIDIA GB200/GB300 rack systems to match GPT-5.4 latency despite significant capability gains. Scores 84.9% on GDPval (44-occupation knowledge work benchmark), 78.7% on OSWorld-Verified (real computer operation), and 90.1% on BrowseComp. API access followed on April 24. Supports 1M token context in the API.

Anthropic ships agent memory in public beta under the Managed Agents API. Agents can now store and retrieve facts, context, and past decisions across sessions — without the developer managing a custom memory layer. Completes the foundational stack for stateful production agents.

OpenAI releases GPT-5.5 and GPT-5.5 Pro, framing it as a step toward a unified AI 'super-app.' Available via API (as chat-latest) from April 24. Positions ChatGPT closer to a full-featured work assistant rather than a pure chat interface.

Microsoft, Google, and Shopify co-publish an open standard for AI agents to discover products and transact on users' behalf, backed by Visa, Mastercard, and 20+ others. First shared infrastructure spec for agentic commerce — defining how agents connect to merchants, authenticate, and complete purchases. Sets the ground rules for a future where AI assistants shop independently.

Google ships the Gemini Enterprise Agent Platform at Google Cloud Next '26: a unified product to build, deploy, govern, and monitor AI agents at scale. Includes an Agent Designer UI, activity inbox, and pre-built agents from Salesforce, ServiceNow, Workday, Adobe, and Atlassian. Paired with a $750M partner fund to accelerate enterprise adoption.

Google confirmed it is Apple's preferred cloud AI partner: Gemini will power the context-aware capabilities in the next major Siri upgrade. Gives Google model reach into every iPhone without a consumer product of its own — and signals that Apple Intelligence's real bet is on partnered foundation models, not Apple-built ones.

Microsoft ships agentic Copilot across all three core Office apps at GA. Multi-step, in-app actions now work natively: Word drafts and restructures documents in one flow; Excel builds analysis, formulas, and charts; PowerPoint updates existing decks with new data while respecting company templates. Takes Copilot from text suggestion to document-level action for all M365 enterprise customers.

OpenAI ships gpt-image-2 with two modes: Instant (fast generation) and Thinking (browses for references, reasons through the prompt, self-verifies before rendering). Significantly better at dense text, precise object placement, and multi-language images. Available in ChatGPT and via API. Thinking mode closes a longstanding gap in instruction-following quality for complex visual tasks.

OpenAI expands Codex from code tool to full agentic workspace: computer use, in-app browsing, image generation via gpt-image-2, memory, and plugins ship in the desktop app. Also launches Codex Labs — a developer training and certification program with GSI partnerships. Signals Codex repositioning as a general agentic surface, not just a coding assistant. Over 4 million weekly developer users.

Anthropic launches Claude Design, an experimental product for creating visual work: prototypes, presentation decks, one-pagers, and UI mockups. Built on Opus 4.7. Users describe what they need, Claude builds a first version, then they refine through conversation or inline edits. Reads a team's design system and codebase for brand consistency. Exports to PDF, URL, PPTX, or Canva. Available to Pro, Max, Team, and Enterprise subscribers.

Opus 4.7 upgrades the flagship model with improved agentic coding, multidisciplinary reasoning, scaled tool use, and computer use. New xhigh effort level gives finer latency vs. reasoning control. Maximum image resolution jumps from 1.15MP to 3.75MP (3.3× increase). Pricing unchanged at $5/$25 per million tokens. Available across Claude products, API, Bedrock, Vertex AI, and Microsoft Foundry.

OpenAI launches GPT-Rosalind, a frontier reasoning model fine-tuned for life sciences: genomics, protein engineering, chemistry, and drug discovery. Named after Rosalind Franklin, the model outperforms GPT-5.4 on six of eleven LABBench2 tasks and leads on BixBench bioinformatics benchmarks. Supports evidence synthesis, hypothesis generation, and experimental planning. Available as a restricted research preview to qualified Enterprise customers; launching with Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific.

OpenAI ships a major Agents SDK update: native sandbox execution isolates agents in scoped environments (only the files, tools, and code each task needs), and a model-native harness optimized for how GPT-5.4 performs on long-horizon, multi-tool work. A Manifest abstraction standardizes workspace description from local prototype to production, with support for S3, GCS, Azure Blob, and Cloudflare R2. Python first; TypeScript planned.

Meta embeds a shopping experience directly into Muse Spark, rolling out across WhatsApp, Instagram, Facebook, and Messenger. Users get outfit suggestions, room styling help, and gift recommendations in-conversation. First major commerce integration built on a foundation model rather than a separate product layer — a template for how AI interfaces could replace traditional e-commerce flows.

OpenAI releases GPT-5.4-Cyber to vetted researchers and security teams, with access tiers that loosen restrictions for verified practitioners doing defensive research. Designed for threat modeling, vulnerability analysis, and exploit documentation — with guardrails calibrated by verification level rather than a single policy. Released in direct response to Anthropic's Mythos Preview from Project Glasswing.

Claude Code's desktop app ships a major redesign: a sessions sidebar for running multiple tasks in parallel, an integrated terminal and file editor with faster diffs, expanded preview pane, side chat (⌘+;) for asking questions without interrupting the main session, drag-and-drop workspace, and SSH support on Mac. The shift from a single-task CLI to a multi-session coding environment.

Claude launches as a native sidebar add-in for Microsoft Word on Mac and Windows. Highlights passages, rewrites sections, and inserts edits as tracked changes using Word's existing review workflow. Completes Claude's integration across the full Office suite (Excel, PowerPoint, Word). Available to Team and Enterprise plans.

Claude Code gains Ultraplan: a cloud-powered planning mode that uses Claude on the web to generate comprehensive implementation plans before coding begins. Designed for complex, multi-file tasks where getting the architecture right up front saves hours of rework.

Anthropic introduces the advisor tool: pair a fast executor model (Sonnet or Haiku) with Opus as a strategic advisor that only gets called on hard decisions. Sonnet + Opus advisor improved SWE-bench Multilingual by 2.7 percentage points while cutting per-task cost by 11.9%. A new paradigm for cost-effective agent intelligence.

Anthropic launches Ask Your Org: a pre-configured Project that searches across connected company tools (Slack, Microsoft 365, Google Workspace, custom MCP connectors) and returns a single synthesized answer with citations. Permission-aware — users only see data they can already access. Available to Team and Enterprise plans after owner setup.

Anthropic ships a batch of admin controls for Team and Enterprise: user groups with SCIM sync, role-based access defining which Claude features each group can use, per-user spend caps, managed Claude Code policies (tool/file/MCP permissions), and a new Compliance API for Enterprise giving programmatic access to usage data and selective deletion.

Cowork goes GA on macOS and Windows with Analytics API access, OpenTelemetry monitoring, and role-based access controls for enterprise departments. The shift from collaborative experiment to production-grade team workspace.

Claude Code gains the Monitor tool: spawn a background process and stream its stdout into the conversation without blocking the thread. Enables patterns like "watch kubectl logs for errors and fix any crashes" — a step toward always-on agent awareness.

Meta launches Muse Spark, its first proprietary (non-open-source) model, developed by Meta Superintelligence Labs. Small and fast, competitive on reasoning and agentic tasks. Signals a strategic shift: Meta now has both open (Llama) and closed model lines.

Anthropic launches Managed Agents in public beta: a fully managed harness for running Claude as an autonomous agent with secure sandboxing, built-in tools, and streaming. Create agents, configure containers, and run sessions entirely through the API. The biggest shift from Claude-as-assistant to Claude-as-worker.

Anthropic announces Project Glasswing alongside AWS, Apple, Google, Microsoft, NVIDIA, and others to secure critical software infrastructure. Claude Mythos Preview — a specialized cybersecurity model — available as a gated research preview for defensive work.

Google quietly ships AI Edge Eloquent, an on-device dictation app for iOS using Gemma-based speech recognition models. Works fully offline; optional Gemini cloud integration for post-processing. Signals Google's push into private, on-device AI — where inference stays on the hardware rather than hitting a remote API.

Google releases Gemma 4, an open model family from 2.3B to 31B parameters that is natively multimodal (text, image, video). The 31B Dense variant ranks #3 globally among open models. A major leap in what open-weights models can do.

Microsoft announces three new proprietary MAI foundation models available in Azure AI Foundry: MAI-Transcribe-1 (state-of-the-art multilingual speech recognition), MAI-Voice-1 (custom voice synthesis), and MAI-Image-2 (top Arena.ai leaderboard scores, 2× faster generation). First clear signal that Microsoft is building its own model stack alongside its OpenAI and Anthropic partnerships.

The 1M token context window is now GA for Claude Opus 4.6 and Sonnet 4.6 at standard pricing — no beta header required. Requests over 200k tokens work automatically. Also raised the media limit from 100 to 600 images or PDF pages per request.

Anthropic launches the Claude Partner Network with $100M invested to accelerate enterprise adoption. Targets system integrators, consultants, and implementation partners helping organizations deploy Claude at scale.

Sonnet 4.6 launches as the balanced model for professional work at scale — improved agentic search, fewer tokens consumed, extended thinking support, and 1M token context window. Web search and code execution tools hit general availability on the same day.

Anthropic raises $30B in Series G funding at a $380B post-money valuation, with $14B annual run-rate revenue. One of the most valuable private companies in history. Signals market conviction that frontier AI will be defining infrastructure.

Opus 4.6 launches for complex, long-horizon agentic tasks. Introduces adaptive thinking (replacing manual budget_tokens), the compaction API for effectively infinite conversations, and data residency controls. Fast mode — up to 2.5x faster — available in preview.

Anthropic launches Labs, a platform for experimental Claude features before they reach the main product. First move toward a formal beta program for early access to capabilities in development.

Anthropic launches HIPAA-ready infrastructure with clinical trial connectors for healthcare and life sciences. Opens Claude to regulated industries that previously had compliance blockers.

Microsoft Copilot reaches general availability across M365 enterprise. AI embedded in Word, Excel, PowerPoint, Outlook, and Teams. Forces the question: if Copilot is good enough for document work, what does Claude add?

Anthropic donates the Model Context Protocol to the newly formed Agentic AI Foundation, cementing MCP as an open industry standard. Adopted by dozens of companies since its release, MCP becomes a neutral, community-governed protocol.

Real-time collaborative sessions in Claude — multiple users working in the same conversation. Foundation for team-based AI workflows.

Opus 4.5 launches as the most capable Claude model yet — step-change improvements in vision, coding, and computer use at a more accessible price than previous Opus models. Best model for complex specialized tasks and professional software engineering.

Claude models come to Azure customers through Microsoft Foundry with Azure billing and OAuth authentication. Full Messages API access including extended thinking, prompt caching, PDF support, and Agent Skills.

OpenAI ships long-term memory as a default ChatGPT feature, and launches canvas — a structured writing and code workspace. Accelerates the shift from chatbot to persistent AI workspace.

Agent Skills launch in beta: pre-built Skills for PowerPoint, Excel, Word, and PDF files. Also supports custom Skills where you package your own domain expertise. Claude can now read and write Office documents without manual parsing — a major unlock for enterprise workflows.

Haiku 4.5 launches as the fastest and most capable Haiku yet — near-frontier performance for real-time applications, high-volume processing, and cost-sensitive deployments where speed matters more than maximum intelligence.

Sonnet 4.5 launches with the highest intelligence of any Sonnet, built for complex agent workflows and coding tasks. Also ships: the memory tool (persistent context across conversations) and context editing for automatic conversation management.

Web fetch tool launches in beta, letting Claude retrieve full content from any web page or PDF by URL. Pairs with web search for end-to-end research tasks.

Claude Code launches as a terminal-native development experience: reads codebases, runs commands, writes and edits files, and navigates full projects. Goes further than autocomplete into agentic development.

Anthropic ships the first Claude 4 models: Opus 4 for frontier tasks and Sonnet 4 for everyday use, both with extended thinking. Also ships: Files API, Code Execution tool, and MCP connector in the API. A step-change in what Claude can do in production.

Web search launches in the API, giving Claude access to up-to-date information. Closes the biggest gap between Claude and web-native AI products.

GPT-4o ships as a single model handling voice, image, and text without switching modes. Real-time voice with sub-second latency. Brings the conversational AI interaction model much closer to reality.

Meta releases LLaMA 4 with performance competitive with GPT-4-class models. Largest open-weights release to date. Reinforces that the gap between open-source and proprietary is closing faster than expected.

Claude Sonnet 3.7 ships as the first Claude model with extended thinking built in — a mode where Claude reasons step-by-step before answering. Near-instant responses or visible thinking, one model. Anthropic's first mainstream reasoning model.

The term "AI agents" crosses from research into mainstream product discussion. Every major AI lab ships or announces agentic products. The question shifts from "can AI do this?" to "how do you orchestrate AI to do this reliably?"

Claude gains the ability to cite its sources when answering from documents — pointing to the exact passage it drew from. Major unlock for trust in enterprise document workflows and RAG applications.

Google ships Deep Research in Gemini Advanced — a mode that autonomously researches a topic across the web over several minutes and returns a structured report. First mainstream implementation of a multi-step research agent.

Anthropic moves several API features to general availability: Message Batches API (50% cost reduction on batch jobs), Token Counting API, Prompt Caching (90% cost reduction), and PDF support. Also ships Go and Java SDKs.

Anthropic releases MCP as an open protocol for connecting AI models to external tools and data sources. Other companies adopt it quickly, making MCP a de-facto standard. The shift from proprietary to standardized AI integrations.

Claude Haiku 3.5 launches as a fast, cost-efficient model for real-time applications and high-volume tasks. Better than Haiku 3 at a similar price point.

Anthropic ships Computer Use in public beta: Claude can move a mouse, click, type, and navigate GUI applications. First mainstream API for AI-controlled computer interaction. Sets the foundation for desktop-level automation.

OpenAI releases o1, a model that reasons explicitly before answering. Significantly outperforms GPT-4 on math and science benchmarks. Introduces "reasoning model" as a distinct category from "chat model."

Workspaces launch in the Developer Console: custom spend limits, grouped API keys, usage tracking by project, and user roles. First real admin layer for teams using the Claude API.

Prompt caching launches in beta: cache and reuse prompt content to reduce costs by up to 90% and latency by up to 80%. Major cost unlock for applications using the same system prompt or documents repeatedly.

Claude Sonnet 3.5 launches with benchmark-leading coding performance. Artifacts ship alongside: a side-panel for code, documents, and interactive outputs. The shift from text responses to living, editable artifacts.

Anthropic ships the Claude 3 model family: three tiers for speed vs. capability tradeoffs. Opus leads on benchmarks. Haiku enables real-time, cost-efficient use cases. Establishes the tiered naming convention still in use today.

Google rebrands Bard to Gemini and ships the Ultra tier. Natively multimodal from architecture rather than retrofit. Signals Google's intent to compete directly with GPT-4 and Claude at the frontier.

GPT-4 launches with vision capabilities and measurably better reasoning than GPT-3.5. Triggers the "AI moment" for enterprise adoption.

Anthropic ships the first public Claude, trained using Constitutional AI — aligning model behavior using a set of principles rather than purely human feedback. First major alternative to GPT at capability scale.

ChatGPT launches and reaches 100 million users in two months — the fastest consumer product adoption in history. Not the most capable model at the time, but the one that made AI accessible to everyone. Everything since has been a response to this moment.