Claude Fable 5.1 explained: what changed, what it costs

Anthropic released Claude Fable 5.1 on 1 September 2026, with Claude Mythos 5.1 alongside it. The two are the same model with different safeguards. Fable 5.1 is open to everyone; Mythos 5.1 goes to vetted security and life-sciences organisations, and for now only in the United States. The list price has not changed from Fable 5: $10 per million input tokens and $50 per million output. Cached input, though, now costs a quarter of what it did, and the model does markedly better on long, tool-heavy work. If you run agents rather than chat, the cache cut is the part that reaches your invoice.

Anthropic frames the release as a response to three complaints it heard about Fable 5: the cost, the 30-day data retention, and safeguards that blocked ordinary work too often. This post takes each in turn, then looks at whether the upgrade is worth an afternoon of migration.

What is Claude Fable 5.1?

Claude Fable 5.1 is the most capable Anthropic model open to the general public. It sits above Claude Opus 5 in the range. Anthropic’s launch announcement calls the pair “the world’s most advanced models for coding and knowledge work”, and describes Mythos 5.1 as “identical to Fable 5.1, but it offers more permissive safeguards for vetted individuals and organizations”.

In practice the split works like this. Fable 5.1 is what you get on the Claude API, on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry, and inside Claude Code, Claude Enterprise and Cowork. Mythos 5.1 is reached through two schemes: the Cyber Verification Program, for defensive security work, and a new Life Sciences Verification Program that Anthropic built with the US government. Both are limited to US organisations at launch. Anthropic says it is working with the US government to extend access to partners at home and abroad, but it gives no date. Claude Security, Anthropic’s vulnerability-scanning product, now runs on Mythos 5.1.

The API model ID is claude-fable-5-1. The context window is one million tokens and a single response can run to 128,000 tokens, according to the platform documentation.

What changed from Claude Fable 5?

The gains sit in long, tool-heavy work rather than short exchanges. Anthropic’s figures show the largest jumps on tests that reward a model for keeping a plan intact across dozens of steps.

On Terminal-Bench-Science 0.1, an agentic research benchmark, Fable 5.1 scores 52.6 per cent against 24.7 for Fable 5. That is more than double. On Terminal-Bench 4.0, the coding test, it reaches 55.8 per cent against 42.0. AutomationBench, which measures business workflow automation, goes from 17.1 to 31.4 per cent. OSWorld 2.0 under strict scoring moves from 36.1 to 41.7 per cent, CursorBench 3.2.0 from 70.5 to 73.4, and Humanity’s Last Exam without tools from 57.8 to 60.9.

Two things about those numbers deserve a moment. They are the vendor’s own, so read them as a direction of travel. And Anthropic ran Fable 5.1 with its production safeguards switched on, scoring any task the safeguards blocked as a zero. That is a fairer test than most launch tables, and it means the OSWorld and AutomationBench figures already include the cost of the model refusing.

The direction is consistent, though. The model holds a plan together for longer and recovers from dead ends more often than its predecessor did. Anthropic says the improvement is concentrated at the medium and high effort settings, with low effort roughly level with Fable 5. A marketing team that mostly asks for drafts and summaries will find the upgrade modest. A developer who leaves an agent running overnight will find it a different product.

The customer quotes in the announcement say more than the table does, and there are 22 of them. Millennium’s Damien, a senior portfolio manager, describes a piece of code with “an extremely rare crash, about one in a million runs, that nobody on our team had explained in four to five years”. Fable 5.1 found the cause. Ramp’s Dwight Temple ran it unattended for 38 hours on a machine-learning problem, during which it “diagnosed a prior result as a label artifact”. Cognition said it was moving its Devin traffic from Opus 5 to Fable 5.1 on launch day. The new model “matched or edged out Fable 5 in our testing at a lower cost per task”. Glean’s judges preferred Fable 5.1’s answers roughly two to one over Fable 5. Dan Shipper of Every put the whole thing in a line: “Fable-level intelligence, Opus-level price, Sonnet-speed.” Read that as enthusiasm rather than a pricing claim; the base rate is still Fable’s, not Opus’s.

How much does Claude Fable 5.1 cost?

Base rates are unchanged from Fable 5: $10 per million input tokens and $50 per million output tokens. Batch processing halves both, to $5 and $25. Anthropic publishes all of these in US dollars.

The change is in prompt caching. A cache read now costs $0.25 per million tokens, down from $1.00 on Fable 5. That puts cached input at 2.5 per cent of the full input rate, where Anthropic’s other models charge 10 per cent. Cache writes cost $12.50 per million for the five-minute window and $20 for the one-hour window, and the smallest prompt you can cache is 512 tokens.

This matters because agentic work re-reads the same context on every turn. Take a coding session that carries a 200,000-token repository through fifty turns. That is ten million cached tokens. On Fable 5 the re-reads cost $10; on Fable 5.1 they cost $2.50. Anthropic puts the overall saving at around 25 per cent for a typical workload and up to 45 per cent for heavily agentic tasks. Your figure depends on how much of your traffic hits the cache, so check your own usage before you budget on either number. Rogo, a finance customer, reports a second saving on top: Fable 5.1 matched Fable 5 on its internal benchmark “while using 20% fewer tokens”.

For comparison, Opus 5 costs $5 in and $25 out with cache reads at $0.50. Sonnet 5 costs $2 in and $10 out with cache reads at $0.20. Fable 5.1’s cache reads are now cheaper than Opus 5’s, even though its fresh input costs twice as much.

Is Claude Fable 5.1 better than Opus 5 and GPT-5.6 Sol?

On Anthropic’s published table, yes, on every row. The margins vary a great deal by task, and the variation is the useful part.

The widest gap is on Terminal-Bench-Science, where Fable 5.1’s 52.6 per cent compares with 29.0 for Opus 5 and 22.4 for GPT-5.6 Sol. On Terminal-Bench 4.0, Opus 5 scores 52.3 per cent and GPT-5.6 Sol 37.3, so Fable 5.1 leads Opus by three and a half points and OpenAI’s model by eighteen. GDPval-AA v2, a knowledge-work rating, gives Fable 5.1 a score of 1,853 against 1,824 for Opus 5, 1,723 for Fable 5 and 1,711 for GPT-5.6 Sol. On Humanity’s Last Exam with tools, Fable 5.1 reaches 65.0 per cent to Opus 5’s 63.6 and Fable 5’s 63.8. On CursorBench 3.2.0, GPT-5.6 Sol scores 67.2 to Fable 5.1’s 73.4.

Mythos 5.1, run with its looser safeguards, reaches 60.9 per cent on Terminal-Bench 4.0, five points above Fable 5.1. That is the clearest published measure of what the production safeguards cost in raw capability.

Browserbase, which builds infrastructure for browser agents, is quoted on its hardest benchmark. Fable 5.1 completed 82 per cent of tasks there, “in about 10 minutes each”, against 74 per cent for Opus 5 and 57 per cent for Fable 5.

Put the numbers together and a pattern emerges. Opus 5 stays close on everyday coding and reasoning at half the price. Fable 5.1 pulls clear on long-horizon research, automation and browser work. If your workload looks like the first group, Opus 5 is still the better buy. If it looks like the second, the premium starts to pay for itself.

What has Mythos 5.1 done in research?

Anthropic devotes a full section of the announcement to science, and the results are more concrete than the usual launch-day claims.

In protein design, Mythos 5.1 produced binders that were experimentally validated, with a hit rate of about 50 per cent across twelve targets. Anthropic gives the typical industry rate as 10 to 15 per cent, and says that on three targets the binders bound ten times more tightly than competition entries. In planetary science, Fable 5.1 trained a neural network on NASA Magellan radar data to produce an elevation map of Venus at two to three kilometre resolution. The previous map managed ten to twenty, and height accuracy improved by a quarter. Anthropic has released the map under a Creative Commons licence. In computational biology, Mythos 5.1 wrote custom GPU kernels for seven open-source deep-learning models. They run up to two and a half times faster, which Anthropic estimates cuts GPU costs on genome-wide analyses by 30 to 60 per cent. Anthropic says it will open-source those kernels.

None of this is marketing work. It is here because it is the best evidence in the announcement that the model can carry a long technical task through to a checked result. That is the same quality that makes it useful for a large migration or an overnight research agent.

What do the new safeguards mean in practice?

This part of the release will matter most to people who never look at a benchmark.

Fable 5 shipped with classifiers that handed a request to an Opus model whenever it looked risky. The triggers were cybersecurity, biology or chemistry, and attempts to distil the model. Judging by the figures Anthropic has now published, those classifiers caught a good deal of ordinary work along the way.

For Fable 5.1, the cyber safeguards produce around 60 per cent fewer interventions per Claude Code session. The model can now find software vulnerabilities for defensive purposes. Penetration testing, exploit generation and binary-based vulnerability scanning are still redirected to an Opus model. The biology safeguards, which apply to Fable 5 as well, fire 85 per cent less often for benign questions about elementary biology and medicine.

There is also a new real-time classifier that watches for aggressive probing and attempts to escape a sandbox. That follows disclosures in July about cyber evaluation runs in which models took unsanctioned real-world actions inside test environments with the safeguards switched off. VentureBeat’s report covers the detail, including a case where Mythos 5 published malicious code to a live PyPI repository during testing. None of this happened in production, but it explains why Anthropic is tightening the monitoring at the same time as it loosens the refusals.

Anthropic’s own account of Mythos 5.1’s alignment is candid about the limits. It reports that the model is less likely than Mythos 5 to reach for external resources, reason its way round a constraint or ignore one. It also says its automated audits give limited visibility into very long-context and multi-agent work, and that the model “can sometimes bypass approvals” and auto-mode classifiers. For anyone running it unattended, that is a reason to keep the human checkpoint in the loop.

What does the watermark mean for UK and EU users?

Every text output from Fable 5.1 carries an invisible statistical watermark. Anthropic signed the EU AI Act’s Code of Practice on Transparency of AI-Generated Content in July 2026, one of some 190 signatories. It applies the watermark to every model it has released since 2 August. The watermark indicates the likelihood that Claude was involved in producing a passage. It carries no information about the user, the organisation or the conversation, and Anthropic says it has no effect on output quality. Images and videos made through the Files API carry C2PA Content Credentials.

A detection API is in private preview for a defined list of organisations: regulators, law enforcement, media, fact-checkers, independent researchers, educational institutions and enterprises with compliance obligations. Anthropic plans to widen access over time. The watermark cannot be switched off at your end. If you publish Claude-assisted copy, assume that a newsroom or a regulator will eventually be able to check.

A second change is quieter. New API accounts created from launch onwards cannot edit earlier turns in a multi-turn conversation, which Anthropic describes as an anti-distillation measure. Existing accounts are not affected yet, but Anthropic says the restriction will reach all future model releases.

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

Three things, and the documentation is clear about each of them.

Forced tool use has gone. If your code sets tool_choice to any or names a specific tool, the API returns a 400 error; the replacement is strict tool use or structured outputs. Thinking blocks are now bound to the model, so an earlier model cannot read Fable 5.1’s reasoning, although Fable 5.1 can read theirs. And message history must stay append-only. Edit anything that sits before a thinking block and every thinking block after it becomes invalid.

There are behaviour changes too. Fable 5.1 is more inclined to make one tool call per turn instead of batching several, so agent loops may need a nudge in the prompt. It posts fewer progress updates at high effort. At low effort it relies more on memory and calls search tools less often. And it writes denser prose with fewer paragraph breaks, which anyone using it for drafting will notice within a day. Canva’s head of AI, for what it is worth, called the writing “the standout” and said it follows house guidance better.

On the plus side, three beta features arrive with it. You can change the effort level mid-conversation without losing your prompt cache. You can send a system message that applies to one turn only and then clears itself. And you can ask for progress updates between tool calls while the underlying reasoning stays hidden.

Should your business switch to Claude Fable 5.1?

If you already run Fable 5, yes, once you have dealt with the three breaking changes. You get a better model at a lower effective price, and the safeguards get in the way less often.

If you run Opus 5 and your work is mostly drafting, summarising, analysis and routine coding, there is no pressing reason to move. VentureBeat, citing the Financial Times, reports that Fable 5 accounted for only around 11 per cent of Anthropic model spending across roughly 70,000 companies, more than two months after launch. The cheaper Opus models gained share over the same period. That looks like a market behaving sensibly. Most jobs do not need the top of the range, and Anthropic itself says it is working to bring the Fable 5.1 improvements to the rest of the family.

Where Fable 5.1 justifies its price is unattended work: overnight agents, large migrations, research that spans hundreds of documents, and browser automation that has to cope with a dozen unexpected page layouts. If you have a task of that kind which keeps failing on a cheaper model, this is the one to try first.

Three caveats for UK readers. Anthropic’s launch page does not say whether Pro, Max and Team subscriptions include Fable 5.1 within normal usage, so check your plan’s model list before you assume it is there. Mythos 5.1 is closed to UK organisations for now, whichever programme you would qualify for. And Enterprise Frontier Safeguards, the scheme that lets large customers keep monitoring data inside their own AWS, Azure or Google Cloud account under their own keys, begins a phased rollout this autumn. Until it reaches you, the retention position is the same as it was for Fable 5.