Claude Fable 5.2: 20% cheaper than 5.1, a fast mode, and three request shapes that now return 400
In brief
On November 9, 2026 Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. Pricing drops from $10/$50 to $8/$40 per million tokens, outputs are up to 30% faster, and a fast mode runs up to 2.5x quicker at double the price. Agentic coding scores jump (FrontierSWE v2: 69.9%, up from 56.3%). Before you switch, check three things that now fail: edited history sent with a later thinking block, the old computer use tool version, and a non-5.2 advisor.
Contents
On November 9, 2026 Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. They are the same underlying model. Fable 5.2 is the version anyone can use, with safeguards that block high-risk cybersecurity and biology work. Mythos 5.2 drops some of those safeguards and is limited to organizations Anthropic has vetted through its Cyber Verification Program or Life Sciences Verification Program.
Fable 5.2 replaces Fable 5.1 as the default Fable model in the Claude apps and Claude Code.
What changed, in numbers
| Fable 5.1 | Fable 5.2 | |
|---|---|---|
| Input price (per million tokens) | $10 | $8 |
| Output price (per million tokens) | $50 | $40 |
| Context window | 1M | 1M |
| Max output | 128k | 128k |
| Output speed | baseline | up to 30% faster |
| Fast mode | no | yes, up to 2.5x faster, $16 / $80 |
| False-positive safeguard blocks | baseline | about 55% fewer |
Anthropic also says Fable 5.2 finishes tasks with fewer tokens at Medium effort, which is the default in the Claude apps and Claude Code. Fewer tokens plus a lower per-token price means the cost per task drops by more than 20% on most workloads.
Benchmarks worth looking at
These are Anthropic's published numbers at each model's highest effort level.
| Benchmark | What it measures | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|---|
| FrontierSWE v2 | Fixing real bugs in code repositories | 69.9% | 56.3% | 62.3% |
| FrontierCode 2.0 | Longer agentic coding tasks | 49.1% | 32.4% | 45.5% |
| Remote Labor Index | Paid freelance projects, end to end | 41.3% | 17.9% | 21.3% |
| OSWorld 2.1 | Operating a computer through the screen | 89.3% | 83.5% | 87.2% |
| AutomationBench | Business workflows across apps | 51.4% | 31.4% | 42.5% |
Two results cut the other way. Gemini 4 Carbon scores higher on AutomationBench (53.9%) and on Chartography, a chart-reading test. And on the AA-Briefcase knowledge-work Elo, Opus 5.5 still edges Fable 5.2 (1800 vs 1785).
The more useful number for budgeting: at High effort on FrontierCode 2.0, Fable 5.2 scores 45.3%, about the same as Opus 5.5's best (45.5%), at roughly half the cost per task. If you moved coding agents to Opus 5.5 for price reasons, rerun that comparison.
Switching: the request shapes that now fail
1. Edited history with a later thinking block
When Claude thinks before answering, the API returns thinking blocks, and you send them back on the next turn so the model keeps its reasoning. Fable 5.2 enforces a prefix check on every account: if your code changed the system prompt, the tools list, or an earlier message, and then sends back a thinking block produced after that point, the request returns a 400 error.
The common way to hit this is a harness that trims or rewrites old tool results to save space, then replays the conversation with the thinking blocks intact. Either stop editing earlier turns, or drop the thinking blocks that came after the edit.
2. The old computer use tool
Fable 5.2 and Mythos 5.2 require the computer_toolset_20260801 toolset. Declaring computer_20251124 returns a 400 error everywhere except Amazon Bedrock. See computer use and browser use for the toolset.
3. The advisor tool
With the advisor tool (beta), a cheaper executor model asks a stronger model for guidance mid-task. If the executor is Fable 5.2 or Mythos 5.2, the advisor must also be Fable 5.2 or Mythos 5.2. Any executor can use Fable 5.2 as its advisor.
A request, with and without fast mode
import anthropic
client = anthropic.Anthropic()
# Standard speed: $8 / $40 per million tokens
response = client.messages.create(
model="claude-fable-5-2",
max_tokens=8000,
output_config={"effort": "medium"},
messages=[{"role": "user", "content": "Review this migration plan and list the steps most likely to fail."}],
)
# Fast mode: up to 2.5x faster output, $16 / $80 per million tokens. No beta header.
fast = client.messages.create(
model="claude-fable-5-2",
max_tokens=8000,
speed="fast",
messages=[{"role": "user", "content": "Summarize the failing test output below in three lines."}],
)
Fast mode doubles the price, so use it where a person is waiting on the answer: interactive coding sessions, live support drafts. Batch jobs and overnight agents should stay at standard speed.
Thinking is always on (adaptive) for both models. You set how hard it works with effort, not with a token budget.
Other things to know before you move traffic
- Mid-conversation switching. On the Claude API and Google Cloud (Google-managed), a conversation that moves onto Fable 5.2 keeps reasoning produced by other models. Moving from Fable 5.2 to any other model drops the Fable 5.2 thinking blocks, and the request still succeeds.
- Thinking blocks are account-bound. Blocks produced by these models only work in the account that produced them, or a linked account. A shared conversation store across two unrelated API accounts will break.
- Data retention. Both models require 30-day retention for safety monitoring and are not available under zero data retention unless Anthropic authorizes it. Anthropic's Enterprise Frontier Safeguards (private beta) runs that monitoring on data kept in your own infrastructure, which is the route for teams that need zero retention.
- Where it runs. Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic-managed and Google-managed), Amazon Bedrock, and Microsoft Foundry. Fast mode is on the Claude API, Claude Code, Google Cloud, and Foundry.
- Blocked requests are answered by an older model. When Fable 5.2's classifiers block a cyber query, Opus 4.8 answers it; blocked biology queries go to Opus 5. Log the
modelfield on responses if you need to know which model actually answered.
For admins and plan owners
Paid subscribers get one to three extra Fable usage resets this week, depending on how long they have subscribed. On Pro and standard Team seats, Fable counts for up to 50% of weekly usage limits. If your team picked models by cost before, choosing the right Claude model walks through the trade-offs; the lower price moves Fable into range for more of that work.
What to do this week
- Search your agent harness for code that edits earlier messages or tool results. Test it against Fable 5.2 before switching.
- Replace
computer_20251124withcomputer_toolset_20260801if you run computer use. - Rerun your cost-per-task comparison against Opus 5.5 at High effort.
- Pick one interactive workflow and try fast mode on it; measure whether people notice.
Announcement: Claude Fable 5.2 and Mythos 5.2. API details: Claude release notes.