Claude Sonnet 5.5: same price, 30% faster, and five breaking changes from Sonnet 5

Anthropic released Claude Sonnet 5.5 on September 28, 2026. It is the second model in the 5.5 family, after Claude Opus 5.5, and the model ID is claude-sonnet-5-5. It is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry.

Anthropic says it produces output more than 30% faster than Sonnet 5 and costs up to 30% less per task, because it needs fewer tokens to finish the same work. Per-token prices did not change.

Price and speed

	Sonnet 5.5
Input	$2 per million tokens
Output	$10 per million tokens
Cache reads	$0.20 per million tokens
Cache writes	$2.50 per million tokens

The saving comes from token count, not the rate card. If your bill is mostly output tokens on long agent runs, that is where you should see it. Measure on your own traffic for a week before you quote a percentage to finance.

What improved

Anthropic's reported numbers:

• Terminal-Bench 4.0, an agentic coding benchmark: 70.6%, against 10.3% for Sonnet 5.
• CursorBench 4.0: 55.5%, against 34.1%.
• GDPval-AA, a test of professional knowledge work: two points below Opus 5.5.

Anthropic positions it for everyday work: bug fixes, documents, summaries, spreadsheets, and design polish, including slide creation that follows a template with minimal editing. Output now carries invisible text watermarking, which Anthropic says supports EU AI Act requirements and does not change how the text reads.

New safeguards

Sonnet 5.5 is the first Sonnet with cyber safeguards and fallback behavior. Prompts that touch sensitive cybersecurity or biology territory can trigger a fallback response, the same approach Opus 5.5 uses. If your team does security research or life-sciences work and sees unexpected fallbacks, that is the cause.

Five breaking changes from Sonnet 5

Code written for Sonnet 5 can fail on Sonnet 5.5 in five ways. Check these before you change the model ID in production.

1. Turning thinking off. thinking: {"type": "disabled"} no longer does it. At high effort or below, send thinking: {"type": "between_tools"} to stop up-front thinking.
2. Forced tool use. tool_choice set to any or tool returns a 400 error. Use auto.
3. Thinking blocks are bound to the model and the conversation. Replaying a block from another context does not work.
4. Computer use. The earlier computer_20251124 tool is rejected on the Claude API and Google Cloud. Amazon Bedrock still accepts it. Move to the current computer use toolset.
5. Advisor tool. It now rejects Claude Opus 4.8, Claude Opus 4.7, and Claude Sonnet 5 as advisor models.

A minimal request that follows the new rules:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=4096,
    thinking={"type": "between_tools"},   # was {"type": "disabled"}
    tools=tools,                          # no tool_choice of "any" or "tool"
    messages=messages,
)

Thinking blocks Sonnet 5.5 produces also work only in the account that produced them, or an account linked to it. If another account sends one back, the API drops the block before the model sees it and the request still succeeds. That matters if you pass conversation history between separate accounts, for example from a vendor's workspace into yours: the model silently loses its earlier reasoning. Blocks from earlier models are not affected. Anthropic's migration guide has before-and-after request examples.

Sonnet 4.5 is deprecated

On September 30, Anthropic deprecated claude-sonnet-4-5-20250929. It retires on November 30, 2026, and Anthropic points to Sonnet 5.5 as the replacement. If any workflow still names it, you have about eight weeks. When your AI model disappears covers how to find those references and test the swap.

What to do this week

• Builders: grep for tool_choice, "disabled", and computer_20251124, then run your evals on claude-sonnet-5-5.
• Admins: find who still calls Sonnet 4.5, including scheduled jobs and no-code tools, and give them the November 30 date.
• Everyone assigning models: use choosing the right Claude model to decide what moves to Sonnet 5.5 and what stays on Opus 5.5. Anthropic says Haiku 5.5 follows in the coming weeks.

Sources: Claude Platform release notes (September 28 and 30, 2026); SiliconANGLE; TestingCatalog.