Claude Fable 5.1 explained: what changed, what it costs

Anthropic released Claude Fable 5.1 on 1 September 2026, alongside Claude Mythos 5.1. They share the same underlying model, with different safeguards and access rules. Fable is available to developers and people on paid Claude plans. Mythos remains restricted to vetted organisations, currently in the United States.

For businesses already using Fable, the immediate change is the cost of reading cached context. The base API price stays at $10 per million input tokens and $50 per million output tokens, but cache reads fall from $1 to $0.25. That matters when an agent repeatedly works through the same documents, instructions or code.

Anthropic presents the release as its answer to three complaints about Fable 5: cost, data retention and safeguards that blocked legitimate work.

Claude model update (checked 24 September 2026)

Anthropic released Claude Opus 5.5 on 22 September 2026 at $4 input and $20 output per million tokens. It now recommends Opus 5.5 as the starting point for most work, with Fable 5.1 kept for demanding reasoning and long-running agentic tasks. The launch figures below are unchanged. Read our Claude Opus 5.5 guide for the full comparison.

The answer in brief

Claude Fable 5.1 is the model Anthropic offers for its most demanding reasoning and long-running agentic work. It was released on 1 September 2026 at $10 input and $50 output per million tokens, and cache reads fell from $1 to $0.25. Since 22 September, Anthropic recommends starting with Claude Opus 5.5 for most work and moving to Fable 5.1 only when Opus 5.5 falls short.

• Best suited to: demanding research, long-running coding, browser work and multi-stage agents.
• API price: $10 input and $50 output per million tokens.
• Cache reads: $0.25 per million tokens, down from $1 on Fable 5.
• API limits: one million tokens of context and up to 128,000 output tokens.
• UK availability: paid Claude plans, the Claude API and supported cloud platforms; included usage depends on your plan.
• Use Opus 5.5 instead when: the work is routine, or Opus 5.5 at a higher effort setting already passes your checks. It costs $4 input and $20 output per million tokens.

What is Claude Fable 5.1?

Anthropic’s model lineup places Fable 5.1 above Opus 5.5 for demanding reasoning and long-horizon agentic work, such as substantial code changes and work involving many documents or tools. The distinction matters most on a task that takes hours and needs repeated checking, rather than a single answer.

Fable includes safeguards that restrict some cybersecurity, biology and chemistry work. Mythos offers more permissive controls to approved users. Its Life Sciences Verification Program is an invite-only beta; Anthropic says Mythos access through the Cyber Verification Program will follow. Claude Security, the company’s vulnerability-scanning product, also runs on Mythos 5.1.

Fable 5.1 is available through Claude on the web, desktop and mobile, Claude Code, Cowork and the Claude API. Anthropic also lists Amazon Bedrock, Google Cloud and Microsoft Foundry. The API documentation gives the model ID as claude-fable-5-1, with a one-million-token context window, up to 128,000 output tokens and a June 2026 knowledge cutoff.

What changed from Claude Fable 5?

Anthropic reports substantial gains on long tasks involving tools. In its launch benchmarks on 1 September, Fable 5.1 rises from 24.7 to 52.6 per cent on Terminal-Bench-Science 0.1 and from 42.0 to 55.8 per cent on Terminal-Bench 4.0. AutomationBench moves from 17.1 to 31.4 per cent, while CursorBench 3.2.0 rises from 70.5 to 73.4 per cent.

The testing used production safeguards, but refusals were handled differently across benchmarks. Blocked tasks received zero for both Fable models on OSWorld 2.0 and for Fable 5 on AutomationBench. Other interventions used Opus fallbacks. Anthropic ran these tests on its own tasks. What they mean for copy quality, search rankings or campaign returns is something only your own trials can show.

Customer accounts give the numbers some context. Millennium reported that Fable 5.1 found the cause of a rare crash that had resisted years of investigation. Ramp described an unattended 38-hour research run. Anthropic picked both for its launch announcement, so read them as ideas for what to test in your own work.

How much does Claude Fable 5.1 cost?

Base rates are unchanged from Fable 5. Batch processing halves the standard input and output rates, to $5 and $25 per million tokens. The figures below are from Anthropic’s API price list, checked on 24 September 2026.

Fable 5.1’s cache reads cost less than Opus 5’s, but Opus 5.5 is cheaper again at $0.20, and its fresh input and output cost 60 per cent less than Fable’s. On a simple request with 20,000 input tokens and 5,000 output tokens, Fable 5.1 comes to $0.45 and Opus 5.5 to $0.18.

A 200,000-token repository read from a valid cache on each of fifty subsequent turns represents ten million cached tokens. The reads alone would cost about $10 on Fable 5, $2.50 on Fable 5.1 and $2.00 on Opus 5.5. That example excludes the initial cache write, newly added tokens and any cache misses, so it illustrates the saving on repeated reads rather than forecasting a complete invoice.

A five-minute cache write costs $12.50 per million tokens; a one-hour write costs $20. Cache expiry and changes to the cached content can add fresh writes. Cloud platform, regional and tool charges may also affect the final bill.

Anthropic estimates that the lower cache rate cuts a typical token-billed Fable workload by about 25 per cent and a heavily agentic one by as much as 45 per cent. Your result will depend on prompt size, cache-hit rate, output length and effort setting. Compare the cost of a finished, checked result, including retries and review time.

Independent testing shows why the headline saving may not reach your invoice. On launch day, Artificial Analysis found that Fable 5.1 at maximum effort used about 1.7 times as many output tokens as Fable 5 on its Intelligence Index, as published on 1 September 2026. Each index task cost $3.76, against $3.14 for Fable 5. The cheaper cache saved about $1.40 per task, but not enough to cover the extra output. Artificial Analysis updates the index regularly, so treat these as launch-day figures.

Effort is the other big lever. Anthropic’s prompting guide says that at medium effort, results roughly match Fable 5 at lower cost. Simon Willison’s launch-day test shows how steep the curve can be: one drawing prompt cost about 10 to 13 cents at low, medium and high effort, $1.83 at xhigh and $3.30 at max. It was one prompt, so treat it as an illustration. It is still a good reason to try lower settings before paying for the highest.

Is Fable 5.1 included in your Claude plan?

UK users can access Fable 5.1, but availability and included usage are different. Anthropic’s Fable plan guide gives these rules:

• Max, premium Team seats and premium seats on legacy seat-based Enterprise plans: up to 50 per cent of the existing weekly allowance can be used on Fable models at no extra charge. The Fable usage comes out of the existing allowance.
• Pro, standard Team seats and standard seats on legacy seat-based Enterprise plans: Fable runs on pay-as-you-go usage credits, subject to the organisation enabling them where required.
• Usage-based Enterprise and the Claude API: usage is billed at standard API rates.

Check your organisation’s model-access settings as well as its spending limits.

How does Fable 5.1 compare with Opus 5.5, Opus 5 and GPT-5.6 Sol?

At launch, the gaps in Anthropic’s comparison varied considerably. On Terminal-Bench-Science, Fable 5.1 scores 52.6 per cent, against 29.0 for Opus 5 and 22.4 for GPT-5.6 Sol. On Terminal-Bench 4.0, the scores are 55.8, 52.3 and 37.3 per cent respectively. Opus is much closer on that coding test. Not every benchmark includes results for every competitor.

Opus 5.5 has since changed the picture. Anthropic says it “performs at the level of Claude Fable 5.1 on most work”. On its Opus 5.5 launch page, Opus 5.5 scores 66.4 per cent on Terminal-Bench 4.0 against 55.8 for Fable 5.1, 40.0 against 31.4 on AutomationBench and 58.7 against 52.6 on Terminal-Bench-Science. The figures come from Anthropic’s own tests, checked on 24 September 2026.

Anthropic’s models overview now says to start with Opus 5.5 for most workloads and to use Fable 5.1 for demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort still falls short. In practice, Fable is worth testing when Opus 5.5 repeatedly stalls, loses track of the task or creates enough rework to wipe out the saving. For more background, see our articles on Claude Opus 5.5, Claude Fable 5 and GPT-5.6 Sol.

What has Anthropic shown in scientific research?

Anthropic’s research examples include protein binders designed by Mythos 5.1 and a new elevation map of a third of Venus, built by Fable 5.1 from NASA Magellan radar data. Both are impressive, but neither tells a business what Fable will do for its own work.

What do the safeguards mean in practice?

Fable 5.1 uses classifiers and fallbacks for work that may involve cybersecurity, biology, chemistry or attempts to extract the model’s capabilities. Anthropic says its revised cyber controls lead to around 60 per cent fewer interventions per Claude Code session than the earlier Fable safeguards. Source-code vulnerability discovery is now allowed, while penetration testing, exploit generation and binary vulnerability scanning are still redirected to Opus models.

In the Claude apps, Anthropic’s help centre says blocked requests currently fall back to Opus 4.8 for offensive cybersecurity and to Opus 5 for biology, chemistry and life sciences. The conversation then stays on Opus until you switch back.

Anthropic also says its newer biology safeguards intervene 85 per cent less often on benign elementary biology and medical questions. Research and development questions in the life sciences can still be redirected. Developers using the API need to configure the relevant fallback behaviour rather than assume every Claude surface handles it identically. Under Anthropic’s refusal rules, a request refused before any output is not billed, but a fallback run is billed at the fallback model’s rates.

Anthropic’s 30 July incident report describes other Claude models reaching real systems during cybersecurity evaluations with safeguards disabled. One run published a malicious package on the live Python Package Index. Both happened during testing, and the report does not describe how deployed Fable 5.1 behaves.

Anthropic’s launch summary of its Mythos 5.1 alignment testing also reports cases of bypassing approvals or automated classifiers. It notes limited audit coverage of very long-context and multi-agent work. If an agent can change data, contact people or affect a live system, keep a human approval step at the point of impact.

What should you check about data retention?

Fable’s retention terms need checking against your account. Under Anthropic’s covered-model policy, affected organisations that would otherwise use zero data retention must allow 30-day safety retention of prompts and outputs, unless they have received an eligibility exception. Consumer plans and other surfaces have their own arrangements. Flagged content or legal requirements can also extend retention.

Enterprise Frontier Safeguards will let eligible customers keep monitoring data in cloud infrastructure they control. The rollout is due to begin in phases in autumn 2026. Eligible customers can receive zero-data-retention access while they wait, so ask your account team whether that exception applies before ruling Fable in or out.

UK teams should also check where the data goes and who keeps it. On Amazon Bedrock, the London region (eu-west-2) supports Fable 5.1 only through global cross-region inference, with no in-region or geographic option, and AWS review of data has to be switched on, according to the Bedrock model card. Third-party tools add their own terms. GitHub Copilot says Fable 5.1 requires data retention by default, with zero-data-retention access for some eligible enterprises until the end of 2026. Vercel’s AI Gateway says prompts and completions are kept for 30 days. Record these routes in your data protection impact assessment before client data goes near them.

What does the text watermark mean for your content?

Fable 5.1 applies a statistical watermark to generated text. It changes how the model chooses between suitable words, without adding hidden characters or identifying the user, organisation or conversation. Anthropic’s marking guidance says it applies worldwide, including the UK.

A detector can estimate whether Claude was involved, but it cannot prove authorship or ownership. Short passages and tightly constrained factual text may provide too little evidence. If Claude only proofreads someone’s writing, its few changes may also be insufficient for detection. Lightly editing Claude-generated text is a different matter: the mark may persist. Anthropic explains these limits in its watermark guide.

Detection is currently in private preview for eligible organisations, including certain public-interest bodies and enterprises with relevant compliance duties.

Supported image, video and audio files can carry separate C2PA Content Credentials when retrieved through the Claude API’s Files API. They record where a file came from. Whether its content is accurate is a separate check.

What should content and marketing teams watch for?

One warning in Anthropic’s notes on Fable 5.1 matters more to a marketing team than any benchmark. When summarising documents, the model is more likely to reproduce passages of the source without marking them as quotations. A research summary or competitor brief can carry someone else’s wording straight into your draft.

Build the check into the workflow. Ask the model to put any wording taken from a source in quotation marks with a link, and run drafts through a plagiarism checker before they are published. Keep a person responsible for every quote and figure that reaches a client or a live page.

What breaks if you move from Fable 5 to Fable 5.1?

For custom integrations using the Messages API, three changes in Anthropic’s migration guide deserve testing:

1. Forced tool choice returns a 400 error. Replace any or tool with auto and explicit instructions. Use strict: true for schema-valid arguments where supported, or structured outputs for JSON. Strict mode checks a call’s arguments but cannot make the model call the tool.
2. Older models lose Fable 5.1’s saved reasoning. The API automatically drops thinking blocks an older fallback model cannot read. The request can succeed, but the model must re-plan.
3. Editing earlier context can cause a 400 error. Preserve the system prompt, tools and earlier turns associated with retained thinking blocks. The documentation dates default enforcement to accounts created on or after 31 August 2026; older accounts can opt in. Use the documented block-dropping option if your workflow needs to discard invalidated reasoning.

People using Claude in the apps have nothing to fix here. Claude Managed Agents handles the migration beyond the model-name change.

Anthropic’s feature notes also describe changes you may notice in daily use. Fable 5.1 can make one tool call where an earlier workflow expected several together, and it gives fewer progress updates during long runs, especially at higher effort. At low effort it searches less often and answers from memory more readily. Its prose can also be denser, with longer sentences and fewer paragraph breaks.

Build them into any trial. A research workflow, for example, should explicitly require fresh sources where they matter. Three beta options offer more control: changing effort during a conversation, instructions that apply to one turn, and readable progress updates between tool calls. Check tool behaviour, fallbacks, caching and total cost before switching a production integration.

Should your business switch to Claude Fable 5.1?

If you already use Fable 5, there is a good case for testing the upgrade. The base rates are unchanged, cache reads are cheaper and Anthropic reports better results on demanding tasks. Custom integrations still need the migration checks above, and the same trial should include Opus 5.5, which may do the job for less.

If Opus 5 or Opus 5.5 already handles your drafting, summaries and routine analysis reliably, Fable 5.1 is unlikely to be worth its higher rates. Keep it for a task where the current workflow costs too much time or repeatedly falls short.

For a marketing team, that could be a competitor brief checked against a known set of sources, or a content-migration audit with a clear list of issues to find. Run the same brief on Opus 5.5 and Fable 5.1. Record the model and tool spend, retries, elapsed time and corrections a person has to make before the work is usable.

That comparison will tell you whether Fable’s rates, two and a half times those of Opus 5.5, buy something your business needs. Anthropic’s own guidance points the same way: start with Opus 5.5, and move to Fable 5.1 for long, multi-step jobs where the cheaper model keeps needing rescue.

Need help choosing or testing an AI workflow?

AIWIZ can help you choose a useful trial, set the acceptance checks and compare the cost of getting the work finished. Talk to us about your AI or automation project.