Claude Fable 5.1: the technical guide to benchmarks, API, pricing and how it compares

01What is Claude Fable 5.1

Claude Fable 5.1 and Claude Mythos 5.1 are the same model; only the level of safeguards differs. Fable 5.1 is available to everyone, while Mythos 5.1 is limited to Anthropic's trusted access programs for cybersecurity and the life sciences. (Anthropic)

When Fable's classifiers flag a request as sensitive (cybersecurity, biology and chemistry, model distillation), a less capable Claude Opus model handles the response instead. (Wikipedia)

A short history of Fable

02The Claude model family in September 2026

Fable 5.1 sits at the top of the public lineup; above it is only the restricted Mythos tier. Below it are Opus 5.5, Sonnet 5 and Haiku 4.5.

Haiku 4.5 is still an active model, but Anthropic lists its retirement as no sooner than October 15, 2026. If you are starting a new project on it, keep an eye on the model deprecations page and plan to move to its successor.

The first Mythos-class model, Claude Mythos Preview, is used by a small group of organizations through Project Glasswing.

03Claude Fable 5.1 technical specifications

Fable 5.1 has a 1 million token context window, up to 128,000 output tokens and adaptive thinking that is always on.

Effort levels and cost

At Low or Medium effort, Fable 5.1 matches or beats Fable 5 at a much lower cost. At higher effort levels it performs far better. (Anthropic)

API change: anti-distillation

New API accounts (created from the day Fable 5.1 launched) can no longer manually edit earlier context in a multi-turn conversation while keeping the record of Claude's thinking. This closes a known technique for extracting the model's reasoning. Existing accounts are not affected yet, but the change will apply to everyone with future models. (Anthropic)

EU AI Act watermark

Models released after August 2, 2026 carry an invisible text watermark. It does not affect output quality and contains no information about the user or the conversation. A detection API is in private preview for regulators, media, researchers and companies that are required to verify watermarks. (Anthropic)

04How to use Claude Fable 5.1

Fable 5.1 has been available across all Claude products and cloud partners since day one.

On Claude.ai and in the apps

1. Open claude.ai or the Claude desktop or mobile app.
2. Select Claude Fable 5.1 in the model picker.
3. The default effort is Medium; raise it for hard tasks.

In Claude Code and Claude Cowork

Claude Code is an agentic tool for developers (terminal, desktop, mobile), and Claude Cowork is its counterpart for knowledge work. Fable 5.1 defaults to High effort in Claude Code and Medium in Cowork. (Anthropic)

Through the Claude API

An illustrative call in Python with the official SDK:

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY from the environment

message = client.messages.create(
    model="claude-fable-5-1",
    max_tokens=4096,
    messages=[
        {"role": "user", "content": "Find the root cause of the failing build in the attached log and propose a fix."}
    ],
)

print(message.content[0].text)

Check the Claude Platform documentation for effort settings, prompt caching and other parameters; names can change.

Cost tip: Cache reads cost just $0.25 per million tokens. Agents that repeatedly read the same context (a repo, documentation, a system prompt) benefit most from prompt caching. (Anthropic)

05Claude Fable 5.1 benchmarks

Fable 5.1's biggest jump is in scientific research: 52.6% on Terminal-Bench-Science, more than double Fable 5. These are Anthropic's numbers, with Fable 5.1 tested with its production safeguards switched on:

Where safeguards stepped in, Fable 5.1 scored zero on OSWorld, and cybersecurity tasks went to Opus 4.8 and biology tasks to Opus 5. The model's real capability is therefore likely higher. Anthropic attributes the gap between Fable and Mythos on Terminal-Bench 4.0 (55.8% vs 60.9%) to older, less precise cyber safeguards. (Anthropic)

Claude Fable 5.1 vs GPT-6 Astra

Both models cost the same in the API ($10 / $50), but each leads somewhere else. On general intelligence they are tied on the current Artificial Analysis index v4.3 (both 53). Fable 5.1 leads on Humanity's Last Exam; Astra leads on computer use, math and terminal tasks. We cover the competitor in detail in our GPT-6 Astra guide. We cover the cheaper alternative from xAI in our Grok 4.7 guide, and even cheaper models for agentic coding in our Gemini 3.8 Flash guide and Muse Spark 1.3 deep dive.

The independent Artificial Analysis Coding Agent Index gives Fable 5.1 70 points against 67 for Astra. (MindStudio)

What early customers report

• Millennium: Fable 5.1 was the first model to find the cause of a rare crash (roughly once in a million runs) that the team had not explained for 4 to 5 years; it disassembled a vendor library and matched it against the core dump.
• Ramp: a single 38 hour unattended run corrected a flawed result, launched six parallel experiments overnight and came back with results and next steps.
• Browserbase: on its hardest browser agent test it completed 82% of tasks, versus 74% for Opus 5 and 57% for Fable 5, while using fewer tokens.
• Every: about 2x faster than Opus 5 with half the tokens.

Source: Anthropic

For all seven September models side by side, with pricing and benchmark charts, see our big comparison of September 2026 AI models.

06Claude Fable 5.1 safety: more precise safeguards

Fable 5.1 raises fewer false alarms than Fable 5 while still protecting sensitive areas. Anthropic documents its testing in detail in the system card.

Cybersecurity

• Fable 5.1 can now be used to find software vulnerabilities, but not to develop exploits.
• Claude Code users will see about 60% fewer cyber safeguard interventions per session on average.
• Penetration testing, exploit generation and binary vulnerability scanning are still routed to Opus models.
• Anthropic had the safeguards tested by two external organizations and automatically through Gray Swan; no critical jailbreak was found.

Biology and chemistry

The new biology safeguards fire 85% less often on harmless questions about basic biology and medicine. Life sciences research and development questions are still handled by Opus models. According to Anthropic's evaluations, Mythos 5.1 does not reach the next risk tier under its Responsible Scaling Policy.

Alignment

Anthropic's automated audit found Mythos 5.1 better aligned than Mythos 5: it is less likely to try to reach outside its test environment, less likely to ignore explicit constraints and less likely to cheat in training. Anthropic also admits the model can occasionally get around approvals and auto mode classifiers. It is Anthropic's most robust model yet against prompt injection on an external benchmark.

Source for this section: Anthropic

The June 2026 export controls

Three days after Fable 5 launched, the U.S. government stepped in. On June 12, 2026, the Department of Commerce sent Anthropic a letter prohibiting access to Mythos 5 and Fable 5 for non-U.S. nationals, so Anthropic suspended access for all customers. (Wikipedia) The Department lifted the controls on June 30, and Anthropic restored access on July 1, 2026. Anthropic's official statement is on its access update page.

Mythos 5.1 for vetted professionals

Mythos 5.1 is available through the Life Sciences Verification Program, built with the U.S. government. The Cyber Verification Program for defensive security currently offers Opus and Sonnet models and is set to add Mythos-class models in the near future. Access is limited to select U.S. organizations for now; Anthropic plans to extend it to international partners. Mythos 5.1 also powers Claude Security, which finds vulnerabilities in code and suggests fixes. (Anthropic)

07Claude Fable 5.1 pricing and availability

Fable 5.1 is on every paid plan, but it is only included in the subscription on Max and on premium Team and Enterprise seats. On Pro and on standard Team and Enterprise seats it is paid separately through usage credits, and usage-based Enterprise plans pay API rates.

Claude.ai plans

Fable models use up weekly limits faster than other Claude models. Using them in Claude Code requires a recent version of the app. (Claude Help Center)

API pricing

The cheaper cache matters most for agents. For typical usage, Fable 5.1 costs about 25% less than Fable 5, and for highly agentic work full of tools and context the savings reach 45%. Anthropic measured this on four weeks of real traffic in August 2026. (Anthropic)

Cloud platforms

Fable 5.1 is available in the Claude API and on Amazon Bedrock, Google Cloud and Microsoft Foundry. (Yahoo Tech)

08Who should use Claude Fable 5.1

Fable 5.1 pays off where mistakes are expensive and tasks are long. For everyday work, cheaper models are enough.

On the Pro plan, Fable 5.1 costs extra, so for routine work it makes sense to stay on Opus 5.5 or Sonnet 5 and switch to Fable only for the hardest tasks. Strengths summarized from Anthropic: Fable.

09How to prompt Claude Fable 5.1

Fable 5.1 works best when you give it a goal, context and a way to verify the result, then let it work.

1. Describe the goal and the definition of done. For example: “build passes, tests pass, no new dependencies.” Fable 5.1 can verify its own work and iterates against that.
2. Give it the full context at once. With a 1 million token window and cheap cache reads, it pays to load the repo, documentation or source material at the start and reuse it.
3. Match effort to the task. Low and Medium are enough for routine work and much cheaper; save High, xhigh and max for hard debugging and research.
4. Write down your style rules. Customers such as Canva report that Fable 5.1 follows writing guidance more closely. (Anthropic)
5. Ask for progress artifacts on long runs. Have it keep notes and finish with a summary plus evidence of what works (tests, screenshots, logs).
6. Frame security tasks defensively. Fable 5.1 will find and fix vulnerabilities, but it will not build exploits.

10Claude Fable 5.1 FAQ

11Verdict: is Claude Fable 5.1 worth it?

Claude Fable 5.1 is a top pick for coding and research, and at launch it led the independent Artificial Analysis Intelligence Index. Today Claude Opus 5.5 is at the top, and Fable 5.1 shares second place with GPT-6 Astra. Compared with Fable 5, it adds performance, more precise safeguards and lower bills thanks to cheaper cache reads.

The weak spots remain the safeguards that hand some cybersecurity and biology tasks to weaker models, and the extra cost on the Pro plan. GPT-6 Astra has the edge on computer use and math. But if you need a model you can trust with a large codebase or an overnight research run, Fable 5.1 is the first choice in September 2026.

Want to put AI agents to work in your own processes? We help you choose the model and connect it to your systems as part of our AI implementation service.

12Sources

• Anthropic: Introducing Claude Fable 5.1 and Claude Mythos 5.1
• Anthropic: Claude Fable
• Anthropic: Fable 5.1 and Mythos 5.1 system card
• Anthropic: restoring access to Fable and Mythos
• Claude Help Center: Fable models on your plan
• Wikipedia: Claude Mythos
• MarkTechPost: release details
• MacRumors: Fable 5.1 at lower cost
• Yahoo Tech: anti-distillation
• Notebookcheck: Fable 5.1 and the Pro plan
• OpenAI: GPT-6 Astra (comparison table)
• MindStudio: GPT-6 Astra vs Fable 5.1 benchmarks
• Claude Platform: model pricing
• Claude Platform: model deprecations
• TeamDay: best AI models of 2026