Claude Fable 5.1
Claude Fable 5.1 is Anthropic's highest-tier model for the hardest reasoning and long agent tasks, but since 22 September 2026 it is no longer the obvious pick. It launched on 1 September 2026, replacing Claude Fable 5 at the same $10/$50 per million token price. Cache reads dropped 75% to $0.25 per million, which Anthropic says cuts typical bills by about 25%.
Fable 5.1 is the same model as Claude Mythos 5.1. The difference is safeguards: Fable has classifiers that refuse some cybersecurity, biology and model-copying requests, and it can hand those off to another Claude model. Three weeks after launch, Claude Opus 5.5 beat it on every test in Anthropic's own table at 40% of the price. Fable 5.1 is still a top model, but most buyers should start with Opus 5.5.
Claude Fable 5.1 is a superb model that got overtaken fast. On 1 September it was the strongest Claude you could buy. It ranked #1 on the Vals Index and more than doubled Fable 5 on Terminal-Bench-Science. The 75% cut in cache-read prices made long agent runs much cheaper.
Then Opus 5.5 arrived on 22 September. In Anthropic's own table it beats Fable 5.1 on all eight tests, for example Terminal-Bench 4.0 at 66.4% vs 55.8%. It costs $4/$20 instead of $10/$50 per million tokens. Artificial Analysis also scores Opus 5.5 higher (58 vs 53 at max effort).
Who should still pick Fable 5.1: teams whose own tests show it wins on their hardest multi-hour tasks, and apps that call it at low or medium effort. There it holds up well against Fable 5 at high effort.
Who should not: anyone watching costs, anyone who needs fast replies, and anyone starting a new project. Try Opus 5.5 at high effort first. Anthropic's own docs now point most workloads there.
Score breakdown
Best for
- The hardest multi-hour agentic coding and research tasks
- Teams whose own evals show Opus 5.5 falls short
- Long agent loops that reuse large cached prompts
What we like
- Top-tier reasoning: #1 on the Vals Index (67.87% without fallbacks) and #1 on LiveCodeBench (90.52%) per Vals AI
- Cache reads cut 75% to $0.25 per 1M tokens; Anthropic says about 25% savings on typical workloads
- Terminal-Bench-Science score more than doubled over Fable 5 (52.6% vs 24.7%)
- Anthropic says cyber safeguards give 60% fewer false positives than before
- Available on every major cloud, unlike Mythos 5.1
Watch out for
- Opus 5.5 now beats it on every test in Anthropic's launch table at 40% of the price
- Slowest model in the current Claude line-up
- Breaking API changes vs Fable 5: forced tool use errors, and editing earlier turns can invalidate thinking
- Safety classifiers can still refuse some legitimate security or biology work
Specs at a glance
| API model ID | claude-fable-5-1 |
|---|---|
| Amazon Bedrock ID | anthropic.claude-fable-5-1 |
| Other platforms | Google Cloud, Microsoft Foundry, Claude Platform on AWS |
| Context window | 1,000,000 tokens |
| Max output | 128K tokens |
| Input / output | Text and images in, text out |
| Thinking | Adaptive, always on |
| Default effort | high |
| Reliable knowledge cutoff | June 2026 |
| Comparative latency | Slower (slowest in the current line-up) |
| Safeguards | Classifiers can refuse cyber, bio/chem and distillation requests; fallback to another model supported |
| Retirement | Not sooner than 1 Sep 2027 |
| Sibling model | Claude Mythos 5.1 (same model, fewer safeguards, invite only) |
Benchmarks
Benchmarks are standard tests. Vendor-run results are marked as such; independent results are preferred where they exist.
| Benchmark | Score | Source | Note |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 55.8% | Anthropic | Fable 5: 42.0%; Opus 5.5: 66.4% |
| Terminal-Bench-Science 0.1 | 52.6% | Anthropic | Fable 5: 24.7%; Opus 5: 29.0% |
| GDPval-AA v2 (knowledge work, Elo) | 1853 | Anthropic | Opus 5: 1824; Fable 5: 1723 |
| OSWorld 2.0 (computer use, partial / strict) | 77.9% / 41.7% | Anthropic | Fable 5: 72.9% / 36.1% |
| Humanity's Last Exam (no tools / with tools) | 60.9% / 65.0% | Anthropic | Fable 5: 57.8% / 63.8% |
| CursorBench 3.2.0 | 73.4% | Anthropic | Fable 5: 70.5% |
| Vals Index | 67.87% | Vals AI | Run without fallbacks; ahead of Opus 5 (67.21%) |
| Artificial Analysis Intelligence Index | 53 (max effort) | Artificial Analysis | Behind Opus 5.5 (58) |
Pricing
| Plan / tier | Price | Notes |
|---|---|---|
| API input | $10 per 1M tokens | |
| API output | $50 per 1M tokens | |
| Cache write (5 min / 1 h) | $12.50 / $20 per 1M tokens | |
| Cache read | $0.25 per 1M tokens | Down from $1 on Fable 5 |
| Batch API | 50% off ($5 / $25) |
Fable 5.1 vs Opus 5.5, which to pick
This is the key question since 22 September. Anthropic's Opus 5.5 launch table compares them directly:
| Test | Fable 5.1 | Opus 5.5 |
|---|---|---|
| Terminal-Bench 4.0 | 55.8% | 66.4% |
| FrontierCode v1.1 | 50.3% | 54.4% |
| CursorBench 4.0 | 51.8% | 57.8% |
| GDPval-AA v2.1 | 1735 | 1846 |
| Humanity's Last Exam | 65.6% | 67.7% |
| OSWorld 2.0 | 80.7% | 81.8% |
| Price (in / out) | $10 / $50 | $4 / $20 |
Anthropic's docs now say to start with Opus 5.5 for most workloads. They suggest Fable 5.1 only when Opus 5.5 at higher effort still falls short in your own tests.
Fable vs Mythos: same model, different safeguards
Anthropic says Fable 5.1 and Mythos 5.1 are the same model. The only difference is how strict the safety filters are.
- Fable 5.1 is open to everyone. Classifiers can refuse some cybersecurity, biology and chemistry, or model-copying (distillation) requests. A refusal comes back as a normal response with
stop_reason: "refusal". Developers can ask the API to retry automatically on another Claude model. - Mythos 5.1 has looser safeguards. It is invite-only, for vetted US cyber defenders and life-science researchers.
Anthropic says the cyber safeguards now give 60% fewer false positives than before.
What changed from Fable 5
Same price, better results, cheaper caching. Developers should know about three breaking changes:
- Forced tool use (
tool_choiceset toanyor a named tool) now returns an error. - Thinking blocks are tied to the model. Other Claude models cannot read Fable 5.1's reasoning blocks, apart from Mythos 5.1.
- Editing earlier turns in a chat can invalidate the saved reasoning. Build your chat history so you only ever add to the end.
New extras include changing effort partway through a conversation, short progress updates between tool calls, and content provenance marking (Anthropic says this supports the EU AI Act).
Launch timeline
- Run your own side-by-side test against
claude-opus-5-5athigheffort before paying Fable prices. On published numbers Opus 5.5 wins at 40% of the cost. - Turn on server-side fallback so refused requests are retried on another Claude model automatically, instead of failing in front of your users.
- Keep chat history append-only. Editing or deleting earlier turns can invalidate Fable 5.1's saved reasoning.
- Lean on prompt caching. At $0.25 per million, cached input costs 2.5% of the normal price, so long agent loops get much cheaper.
- Try
lowormediumeffort for routine steps. Anthropic says Fable 5.1 at those levels matches or beats Fable 5 at higher effort.
Jargon explained
- Safety classifier
- A separate filter that checks requests and blocks certain risky topics, such as building cyberattacks or dangerous chemicals.
- Fallback
- When one model refuses a request, the system automatically retries it on a different model.
- Distillation
- Training a new AI model by copying the outputs of an existing one. AI companies try to block it to protect their models.
- Humanity's Last Exam
- A very hard test of expert-level questions across many subjects, built so that top AI models still struggle with it.
- Vals Index
- An independent score from Vals AI that combines results from many real-world tasks, such as legal, finance and coding work.
Alternatives to consider
Claude Opus 5.5
Higher scores on every test in Anthropic's table, faster, and $4/$20 per 1M tokens. The default pick for most teams.
Claude Fable 5
The predecessor at the same price but with 4x pricier cache reads. Only for apps that depend on forced tool use.
Claude Mythos 5.1
The same model with looser safeguards, for vetted cyber-defence and life-science teams.
GPT-6 Astra
OpenAI's top rival; scored 57.9% on Terminal-Bench 4.0 in Anthropic's own comparison.
Gemini 3.5 Pro
Google's announced flagship; still unreleased (partner testing only) as of September 2026.
Frequently asked questions
When was Claude Fable 5.1 released?
On 1 September 2026, together with Claude Mythos 5.1.
How much does Claude Fable 5.1 cost?
$10 per million input tokens and $50 per million output tokens, the same as Fable 5. Cache reads cost $0.25 per million, 75% less than Fable 5, and the Batch API halves prices.
Is Fable 5.1 better than Opus 5.5?
No, not on published benchmarks. Opus 5.5 beats Fable 5.1 on all eight tests in Anthropic's launch table and on the Artificial Analysis index, at 40% of the price. Fable 5.1 may still win on some very long, hard tasks, so test both.
What is the difference between Fable 5.1 and Mythos 5.1?
Anthropic says they are the same model. Fable 5.1 has stricter safety classifiers and is open to everyone. Mythos 5.1 has looser safeguards and is only for approved Project Glasswing and trusted-access members.
Why did Fable 5.1 refuse my request?
Fable's classifiers refuse some requests about cybersecurity, biology and chemistry, or copying the model. The API returns stop_reason: "refusal". You are not billed if it refuses before writing anything, and you can set up automatic fallback to another Claude model.
Can I use Fable 5.1 in the Claude app?
Yes. Fable models are offered on Anthropic's higher-tier Claude subscription plans, as well as the API and major clouds.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- Introducing Claude Fable 5.1 and Claude Mythos 5.1 (Anthropic)
- Claude Fable 5.1 model overview (Claude Platform Docs)
- Introducing Claude Opus 5.5 (Anthropic)
- Models overview (Claude Platform Docs)
- Anthropic's new Fable release is cheaper, less restrictive (TechCrunch)
- Vals AI: Claude Fable 5.1 (Vals AI)
- LLM Leaderboard: Intelligence Index (Artificial Analysis)
- Claude Mythos (Wikipedia)