Claude Fable 5.1: built to work for hours, and told you not to buy it

Key facts

• Headline price $10 in / $50 out per million tokens; cache reads $0.25 (checked 6 Sept 2026) Source: Anthropic pricing page
• Independent score Intelligence Index 57 — first of 202 models — but $6.12 per task against $4.21 for Opus 5 Source: Artificial Analysis
• Its maker's advice Start with the cheaper model; move up only if your own evaluations say so Source: Claude documentation

What it is actually for

Most AI you have used answers a question and stops. This one is aimed at something else: a job you hand over and come back to.

Anthropic markets it as the model for "the hardest knowledge work and coding problems" — ambitious, long-running projects that span several applications. The industry word for this is agent: software that takes a goal, plans its own steps, uses whatever tools it needs, notices when a step fails and tries again, without checking in. The technical shape of that promise is in the documentation rather than the marketing: Anthropic lists the gains as concentrating in six areas — agentic coding over long sessions, document and spreadsheet and slide work, multi-step research, reading dense charts and tables in PDFs, reasoning across the full 1 million token context window, and operating a browser or desktop application and recovering from failed steps. A context window is the amount of material the model can hold in mind at once; a token is roughly three quarters of an English word, so a page of text is about 400 tokens.

That is a real category change, and it comes with a real trade. Something that works unsupervised for three hours can also be wrong unsupervised for three hours. Which brings us to the improvement Anthropic put first.

Quick reference

Everything in this table comes from Anthropic's developer documentation rather than its launch post, which matters: these are contractual, checkable numbers, not marketing. All prices are from Anthropic's pricing page, checked 6 September 2026 and converted at approximately 1:1 EUR/USD for readability; your invoice is in dollars.

The improvement that matters most is not a score

Anthropic's headline claim is not raw ability. It is better judgement on ambiguous work, and fewer confident wrong answers.

If you have worked with these tools you already know why that leads. A model that says "I am not sure, here are two readings of your request" costs you a minute. A model that picks one interpretation, sounds completely certain, and builds three hours of work on top of it costs you a week — and it costs you the trust of whoever you showed the work to.

The technical support for that claim is thinner than the marketing, and the honest place to find it is Anthropic's own system card — the document a lab publishes describing what it tested and what it found, including the bad news. On the automated behavioural audit, Anthropic writes that the model "is a slight regression on overall misaligned behavior compared to Opus 5", and that it "cooperates with human misuse and accepts unverifiable claims of authorization somewhat more readily than Opus 5" (system card, page 3), while being less likely to ignore explicit constraints, hallucinate inputs, or falsely claim to have finished a task. So: better at not fabricating, slightly worse at pushing back.

What this means for you. Confident wrongness is the expensive failure in this whole field, and the vendor's own audit says this model is not uniformly better at it. Put a line in your test plan that counts confident mistakes separately from uncertain ones, and a second line that counts how often the model does something because a user asserted they were allowed to.

The scores, in plain language

A benchmark is a standard exam every company's model sits so results can be compared. Every figure in the table below is as claimed by Anthropic in its own launch post; none of them has been independently reproduced at the time of writing. The useful thing here is not the numbers but where the gaps are.

Read it by the size of the jumps, not the size of the numbers. The first three roughly doubled or came close; those are all the same kind of test — many steps, real tools, recover from failure. The last four moved a few points, 2.9 points on CursorBench and 3.1 on Humanity's Last Exam by Anthropic's own figures. That tells you precisely where the generation improved, and it is not "answers questions better".

The one independent measurement we could find

Artificial Analysis runs its own harness rather than reprinting a vendor's slide, and its results are the counterweight this launch otherwise lacks. It places Fable 5.1 first of 202 models on its Intelligence Index with a score of 57, against 54 for Claude Opus 5, in fourth place. Three points of index, for double the token price. It also measures the boring things vendors omit: Fable 5.1 takes 264.91 seconds to produce its first token and costs $6.12 per Intelligence Index task, against 62.06 seconds and $4.21 for Opus 5.

What this means for you. Independently measured by Artificial Analysis, the top model is about four and a half minutes slower to say its first word and roughly 45% more expensive per task, for three points of composite score. If a human being is sitting there waiting, that is not a close decision.

What is missing from the table

Anthropic published its number on the real-world-job benchmark — GDPval-AA v2, 1853 against 1723 — which is the closest thing the industry has to a test about doing work rather than passing exams. When OpenAI launched GPT-6 Astra two days later, that category of test was absent from the announcement, and Astra reportedly does worse on it. We mark that last clause UNVERIFIED: the Hebrew coverage that carried it, Geektime, returned a 403 error to us on 6 September 2026, both through our fetching tool and through a browser user agent. The habit is worth more than the claim: check what is missing from a comparison table, not only what is in it.

Three things it actually did

Benchmarks are abstractions. These are the concrete results Anthropic put forward, and they are more informative than any percentage — partly because you can picture them, and partly because you can see what kind of work this is good for. All three are the company's own examples, chosen at launch.

The Venus map is the one to hold on to, because it is the clearest illustration of where this technology pays. The data had been sitting in an archive for decades. Nothing was missing except the capacity to work through it properly. That is a description of a great deal of work in a great many organisations — the analysis nobody got round to, the archive nobody re-read, the backlog that was never worth a person's month.

Two cautions. First, two of the three headline results were produced with Claude Mythos 5.1, the restricted configuration you cannot buy. Second, the customer numbers in the same announcement are testimonials, not reproductions: Browserbase reports 82% of tasks completed against 74% for Opus 5 and 57% for Fable 5, and Rogo reports matching Fable 5's accuracy with 20% fewer tokens, both supplied as part of the launch. Treat all of it as evidence of what the tool is for, not as results you can bank.

Why the price tag stopped telling you anything

The price per page of text did not change. Something else did, and it is the actual release.

When a model works on a long job, it re-reads the same background material at every step — the same documents, the same project, the same history — with only a little new information each time. Providers charge less for material the model has already seen; this is called prompt caching, and you can think of it as a returning-customer rate. There are three prices, not one: a premium to put material into the cache the first time, a cheap rate to read it back, and the normal rate for anything new.

What changed is the middle one. A cache read on Fable 5.1 costs $0.25 per million tokens against $1 on Fable 5 — a 75% cut, and 2.5% of the base input price where every other Claude model charges 10% (checked 6 September 2026). Anthropic's estimate of the effect is "around 25% relative to Fable 5" for typical workloads and "up to around 45%" for complex coding and highly agentic tasks. VentureBeat restates the same two figures and labels the accompanying benchmarks "vendor-reported results rather than independent proof of superiority". Nothing about the headline price moved.

What this means for you. Two consequences, neither technical. First, compare cost per finished job, not cost per page: a model that costs twice as much and needs half as many attempts is cheaper, and a price-per-page comparison will confidently tell you the opposite. Second, ask whether your setup actually earns the discount — it depends on how the work is organised, not on which plan you bought, and the 512-token minimum means a lot of chat-shaped usage never qualifies.

What three real jobs cost

Per-token rates are not pricing. Below are three jobs with a token estimate and a euro figure, all computed from Anthropic's published rates at approximately 1:1 EUR/USD, checked 6 September 2026. The arithmetic is ours and is shown so you can redo it with your own numbers.

The third row is the interesting one, and it is not a typo. On the published rates, that run costs 2.50 EUR of cache writes plus 5.00 EUR of cache reads plus 7.50 EUR of output on Fable 5.1, and 1.25 plus 10.00 plus 3.75 on Opus 5. Both total 15.00 EUR. The same run on the old Fable 5 cache rate of 1 EUR per million would have cost 30.00 EUR, which is where Anthropic's "up to around 45%" comes from — our version of that workload lands at 50%, and the exact figure depends entirely on the ratio of re-reading to writing.

There is a rule hiding in that arithmetic, and it is the most useful thing in this article. Against Opus 5, Fable 5.1 is dearer on every line except cache reads: double on input, double on output, double on cache writes, half on cache reads. So Fable 5.1 only undercuts Opus 5 when the cache reads outnumber the output tokens by roughly a hundred to one, plus about twenty-five times the cache writes. Below that ratio, the model marketed as the economical choice for long agent runs is the more expensive bill.

What this means for you. Before anyone signs anything, pull one week of real usage and get the four numbers: fresh input tokens, cache write tokens, cache read tokens, output tokens. If your cache reads are not at least a hundred times your output, the cheaper model is also the cheaper model. And if nobody is waiting for the answer, the Batch API halves both rates, which is a larger saving than the entire cache change.

Which of the two do you need?

Anthropic sells Opus 5 at half the price — 5 EUR in and 25 EUR out per million tokens against 10 and 50. On several everyday measures the gap between them is a few points. Here is how to tell them apart without a spreadsheet.

Note the last row but one. Fast mode — Anthropic's premium low-latency tier — is offered on Opus 5 and Opus 4.8, not on Fable 5.1, and at that premium Opus 5 costs exactly what Fable 5.1 costs at standard speed. If your problem is that answers take too long, the top model is not the fix; it is the opposite of the fix.

What this means for you. For most teams, most of the time, the cheaper model is the right answer, and the top model is a targeted upgrade for one identifiable kind of work: long, unattended, tool-heavy, nobody waiting. Everything else — chat, drafting, review with a person in the loop, anything latency-sensitive — points the other way, and does so on price, on speed and on the vendor's own advice at once.

The quiet fix nobody will write about

Anthropic reports that the model now wrongly refuses far fewer reasonable requests. Two figures, both from the launch post and both the company's own: "Claude Code users can expect an average of around 60% fewer interventions per session from our cyber safeguards, relative to the previous safeguards on Fable 5", and the biology safeguards "fire 85% less often for benign requests related to elementary biology and medical questions".

If you have never had an AI refuse to answer a perfectly normal question, this sounds like housekeeping. If you have, you know it is not. Over-refusal is not a safety feature, it is a reliability failure: it teaches people that the tool is unreliable, and they stop using it for the things it would have been genuinely good at. This kind of improvement never makes a headline and often decides whether a rollout survives its first month.

What happens when a safeguard does fire

This is the mechanism the launch post does not explain, and it is worth understanding before you buy. Fable 5.1 runs behind blocking classifiers for cybersecurity and biology. When one fires, the request does not simply fail — it is answered by a different, older model. Anthropic documents the permitted fallback targets as Claude Opus 4.8 and Claude Opus 5, and refunds the prompt-cache cost of the switch. You are not billed for a refusal that arrives before any output.

The rates are not small. On the external prompt-injection benchmark Anthropic reports, "roughly half of Fable 5.1's coding rollouts fell back to Opus 4.8, compared to under 10% in computer use and tool use, for an overall fallback rate of 23%" (system card, page 83). Anthropic is also candid that the trade is deliberate: because of the model's increased cyber capabilities it has "opted for a wider safety margin", so the classifiers "will continue to block some benign or borderline uses out of an abundance of caution" and are "still likelier to trigger than Opus 5's safeguards" (page 3).

What this means for you. On security-adjacent coding work you may be paying the top model's price for an older model's answers a large fraction of the time. That is not hidden — it is in the system card — but it will not appear on your invoice as a separate line. If your team writes or reviews security code, measure how often stop_details reports a classifier fired before you decide the upgrade is worth it.

Prompt injection: two benchmarks, two answers

Prompt injection is an attack where instructions are hidden in something the model reads — a web page, a document, a ticket — and the model follows them. It is the main reason to be careful about letting an agent loose on your systems, and the two measurements Anthropic publishes disagree with each other. On the Gray Swan indirect-prompt-injection benchmark run with the UK AI Security Institute, Fable 5.1 is the most robust Claude yet at 0.1% attack success after one attempt and 1.0% after fifteen, against 0.4% and 4.8% for Opus 5; the strongest non-Claude model tested reached 9.2% and most sat between 24% and 53% (page 83). On Gray Swan's adaptive Shade tool, which lets an attacker refine against the same scenarios 200 times each, Fable 5.1's attack success rate in coding environments is 9.26% against 2.68% for Opus 5 and 0.15% for Sonnet 5 (page 85) — and the gap is "almost entirely explained by the cyber classifier fallback", since 95% of the requests it answered were served by Opus 4.8.

Both numbers are real and they measure different threats: a library of transferable attacks, and an adversary who gets to practise on you. Report both to whoever owns your security posture. Anthropic's own summary of its residual risk is the honest note to end on: "On alignment risks, we now assess the risk of catastrophic harm as low rather than very low" (page 2). Design as though the model will occasionally do the wrong thing, and give it the smallest set of systems that lets it finish the job.

The version you cannot buy

Alongside Fable 5.1, Anthropic released Claude Mythos 5.1. It is the same model with more permissive safeguards in biology and cybersecurity — "same capabilities as Claude Fable 5.1", identical specifications and identical pricing, available to Project Glasswing participants only. You cannot buy it at any price; access runs through your Anthropic, AWS or Google Cloud account team.

Project Glasswing began on 7 April 2026 as a defensive-security programme with 11 launch partners including AWS, Apple, Microsoft, Google and JPMorganChase, plus more than 40 further organisations that maintain critical software infrastructure, backed by $100M in usage credits and $4M in donations to open-source security work. On the life-sciences side, Anthropic says that "in partnership with the US government, we have enrolled our first participants" and that it is "coordinating with the US government to expand access to a broader set of domestic and international partners". Hand-picked, in other words, and currently US-centred, but not permanently closed.

Two days later OpenAI did the same thing with Astra: full capability for vetted organisations, a restricted version for the market. Two competitors, one week, the same structure. That is now the industry's standard answer to its most dangerous capabilities.

What this means for you. The model in the impressive headline and the model you can license are increasingly not the same thing. Two of the three flagship results in this launch — the protein design at nearly 50% hit rate and the GPU work at up to 2.5 times faster — were done with Mythos. Read those as a ceiling, not an offer.

Where you actually meet it

Most people will never touch the API. Anthropic lists Fable 5.1 for Pro, Max, Team and Enterprise users, and it turns up in several products at once.

• In the Claude apps, on the paid plans, and in Claude Code, the version developers run in a terminal (Anthropic).
• In Cowork, working through a backlog of tasks, and in Slack through Claude Tag, currently in beta (Anthropic).
• Through the big clouds. AWS says there are two routes on its side alone — Amazon Bedrock and Claude Platform on AWS, both from 1 September 2026. Google Cloud and Microsoft Foundry carry it as claude-fable-5-1 too.

Two settings deserve an owner. First, how hard the model thinks is adjustable through an effort parameter, thinking is always on and cannot be disabled, the default is high, and Anthropic's own migration advice is to "re-tune effort from the default". Thinking is billed like output, so a default nobody revisited is a standing charge. Second, per-message effort is now in beta, which lets you raise the level for one hard step and drop it again without invalidating the prompt cache — that is the difference between a targeted splurge and paying premium rates for a whole session.

If you are migrating from Fable 5, three changes will break working code rather than merely degrade it: forced tool use now returns a 400 error, earlier models cannot read this model's reasoning blocks, and editing an earlier turn in a conversation invalidates them. Anthropic also discloses behaviour changes that cost money without breaking anything — it may issue one tool call per turn where Fable 5 batched several, and it is more likely to rewrite a whole file for a small edit.

Prompt templates

These are starting points for the three jobs this model is actually good at, written against the behaviour Anthropic documents for Fable 5.1 — in particular that it batches tool calls less eagerly than its predecessor, writes fewer progress updates, and is more likely to reproduce source passages without marking them as quotations. Each template compensates for one of those.

Template 1: a long unattended run, with a stop condition

Goal: [one sentence, with the finished artefact named].

Before you start, restate the goal in your own words and list the assumptions
you are making. If any assumption would change the result, stop and ask.

While working:
- Batch independent reads into a single turn wherever the next steps do not
  depend on each other.
- Write one short progress line before each tool call: what you found, what
  you will do next.
- When you edit a file, make a targeted edit. Do not rewrite the whole file.

Stop and report if: you have made three consecutive failed attempts at the
same step, or the work exceeds [N] tool calls, or you need a credential.

Deliver: [artefact], plus a list of everything you were unsure about.

Template 2: document review that does not quietly plagiarise the source

Review the attached document against the standards in the system prompt.

Produce:
1. Key obligations for each party, max 300 words.
2. A table of deadlines and notice periods, with clause numbers.
3. Anything that conflicts with our standards, with the clause number.
4. Anything unusual or missing versus a standard German commercial contract.

Rules:
- Any text taken from the document must be in quotation marks with its clause
  number. Everything else must be your own words.
- Do not invent clauses. If a point is ambiguous, say so and quote the text
  that makes it ambiguous.
- Separate "I am confident" findings from "worth a lawyer's eye" findings.

Output language: English.

Template 3: the eval that decides which model you buy

[Your real task, written exactly as your team would write it.]

At the end of your answer, and separately from it, output:
- CONFIDENCE: high / medium / low
- ASSUMPTIONS: every assumption you made that the task did not state
- UNVERIFIED: every claim you could not check against a source you were given

Do not soften the answer. If the task cannot be done as specified, say so.

The third one is the important one. Run it unchanged on both models, on the same three real tasks, and score the answers blind. Anthropic's own instruction is to move up only "when your evals on Claude Opus 5 at higher effort still fall short", and this is the smallest thing that counts as such an eval.

Suited and not suited

Not every task benefits from a model this expensive and this slow. The split below follows the capability list Anthropic publishes and the measured latency and cost from Artificial Analysis.

What to settle before you roll it out

These are questions for the people who own them in your company — your data-protection officer, your legal counsel, your works council — not legal advice. But the first two have documented answers that will surprise people, so start there.

1. Find out where inference actually runs. On Anthropic's first-party API the inference_geo parameter accepts only us or global, workspace storage geo is currently only us, and pinning to the US costs 1.1x on every token category. There is no EU-only option there. If EU processing is a requirement, the question becomes which cloud route — Amazon Bedrock, or Google Cloud, both of which sell regional endpoints — and whether your data-protection owner will sign it.
2. Ask about data retention before the first real prompt. Anthropic states that Fable 5.1 and Mythos 5.1 "carry 30-day data retention and aren't available under zero data retention unless expressly authorized by Anthropic". If your existing arrangement assumes zero retention, this model is not covered by it until someone renegotiates.
3. Watch the "on your own cloud" option, but do not plan on it yet. Anthropic describes an enterprise arrangement in which "customers store their data on their own cloud infrastructure, rather than on Anthropic's systems", with human review done by the customer, rolling out in phases from autumn 2026. That is a roadmap item, which makes it a question about your timeline rather than your contract.
4. Sign the Auftragsverarbeitungsvertrag before the pilot, with whichever party is actually processing — Anthropic, AWS, Google or Microsoft — and get in writing what is and is not used for training.
5. Bring in the works council early if this changes how people work. Under the Betriebsverfassungsgesetz, software that monitors or reshapes work is normally subject to co-determination. A defined pilot is far easier to agree than a rollout that arrives as a surprise.
6. Ask whether AI-generated output needs labelling in your context. Anthropic applies a statistical text watermark to all Fable 5.1 output and signed C2PA Content Credentials to media retrieved through the Files API. That is a provenance mechanism, not a compliance answer, but it is a useful fact for whoever is reading the EU AI Act's transparency duties on your behalf.
7. Fix the model version in writing. Anthropic commits to not retiring claude-fable-5-1 before 1 September 2027, and the clouds set their own dates. If a test you run in September justifies a decision in December, it has to be repeatable against the same version.
8. Give anything unattended the narrowest access that works. Improved injection resistance lowers the odds; it is not a wall, and Anthropic's own alignment risk assessment moved from "very low" to "low" this generation.

Two weeks to a decision

A pilot answers the question no benchmark can: does this save time on our tasks, with our systems, at a cost we can justify? Two weeks is enough, because the decision Anthropic's documentation asks you to make is binary — Opus 5, or Fable 5.1 where Opus 5 falls short.

1. Days 1 to 2: pick three real tasks with a judgeable outcome, at least one of which genuinely runs long — something that takes an hour and touches several tools.
2. Days 3 to 4: agree what "done properly" means before you touch the model. Who judges it, and how many corrections are acceptable.
3. Days 5 to 7: run all three on Opus 5 first, using the eval template above. This is your baseline, and for a good share of teams it is also your answer.
4. Days 8 to 10: run the same three on Fable 5.1. Record the cost of a finished job including failed attempts, and count confident mistakes separately.
5. Days 11 to 12: pull the four token numbers from your usage — fresh input, cache writes, cache reads, output — and apply the ratio test from the pricing section. This step alone sometimes reverses the ranking.
6. Days 13 to 14: decide task by task. Some work stays on the cheaper model, some moves. Writing down which and why is the deliverable.

One related note: if a model is going to work unsupervised for hours, the plan you approve at the start is the last cheap place to catch a wrong assumption. We wrote separately about how the instructions you give these tools should change as they get smarter — it turns out that over-explaining now makes the results worse.

Bottom line

This is a real step forward in a narrow and valuable place: work that runs long, uses many tools, and has to recover from its own mistakes. A doubling on Terminal-Bench-Science and AutomationBench is not marketing noise, and the cheaper re-reading changes the economics of long jobs rather than merely offering a discount.

But three things in this release point the same way, and none of them is in the headline. Anthropic tells you to start with Opus 5. The one independent measurement we found puts the gap at three index points for 45% more cost per task and four times the wait. And the arithmetic on the published rates says the top model only wins on price when re-reading swamps everything else.

Take the advice. Decide what "good" means, run the cheap option first, and let a documented failure be the thing that moves you up. That order costs two days and can halve a bill. The other order costs a quarter before anyone notices.

What this guide did not check: we could not read the Hebrew coverage at Geektime — it returned 403 to both our fetching tool and a browser user agent on 6 September 2026 — so the claim that GPT-6 Astra scores worse on the real-world-job benchmark stays UNVERIFIED. Every capability and safety score here except the Artificial Analysis figures is Anthropic's own, published by Anthropic, and none has been independently reproduced. We did not test the model ourselves, and we did not verify the customer testimonials.

FAQ

Official sources

• platform.claude.comAnthropic: Claude Fable 5.1 model page — model IDs, context window, prices, availability, retirement date
• platform.claude.comAnthropic: What's new in Claude Fable 5.1 — breaking changes, effort, cache pricing, fallback, retention, provenance
• platform.claude.comAnthropic: Pricing — per-model rates, cache multipliers, batch discount, data-residency multiplier
• platform.claude.comAnthropic: Models overview — the "start with Claude Opus 5" guidance, cutoffs and latency classes
• platform.claude.comAnthropic: Data residency — inference geo values, workspace geo, and the 1.1x US multiplier
• anthropic.comAnthropic: "Introducing Claude Fable 5.1 and Claude Mythos 5.1" — benchmark claims, cost estimates, safeguard figures, customer testimonials
• anthropic.comAnthropic: Claude Fable 5.1 and Mythos 5.1 system card (PDF, 212 pages) — prompt-injection results, classifier fallback rates, disclosed regressions
• anthropic.comAnthropic: Claude Fable product page — plans, Claude Code, Cowork and Slack availability
• anthropic.comAnthropic: Project Glasswing — the vetted-access programme that gates Claude Mythos
• artificialanalysis.aiArtificial Analysis: Claude Fable 5.1 — independently measured intelligence index, throughput, latency and cost per task
• artificialanalysis.aiArtificial Analysis: Claude Opus 5 — the same measurements for the cheaper model
• aws.amazon.comAWS: Claude Fable 5.1 is now available on AWS — Bedrock and Claude Platform on AWS routes
• venturebeat.comVentureBeat: Fable 5.1 and Mythos 5.1 arrive with a 75% cost reduction for cache reads
• geektime.co.ilGeektime: Anthropic launches Fable 5.1 and Mythos 5.1 (Hebrew; returned 403 to us on 6 September 2026, so nothing here rests on it)