Anthropic released Claude Fable 5.2: 20% cheaper and up to 30% faster. What the measurements show

In short

Every chart is labelled with whose numbers it shows: Anthropic’s own or an independent measurement. Where I did the arithmetic myself, I say so.

01

What was released

Anthropic released one new model in two configurations. Claude Fable 5.2 is generally available and has safeguards that stop it from doing dangerous work in cybersecurity and biology. Claude Mythos 5.2 is the same model with looser restrictions in those areas.

The split dates from the spring. According to CNBC, Mythos was introduced in April with restricted access because of its hacking abilities, and Fable followed two months later as a public version with extra safeguards. Version 5.1 came out in September. I wrote about the first release separately.

What is known about the model itself:

• the API identifier is claude-fable-5-2;
• the context window is 1 million tokens, and a reply can run to 128 thousand tokens;
• its knowledge runs to June 2026;
• it outputs text only: it reads images but does not make them;
• Fable 5.2 is available on all paid Claude plans, in Claude Code, on the Claude Platform, and on Amazon Web Services, Google Cloud and Microsoft Azure;
• Mythos 5.2 is granted on application through two verification programs, one for cybersecurity professionals and one for life sciences researchers. Enterprise customers get it as part of the Claude Security product.

Anthropic is not telling everyone to switch to Fable. The announcement says that for most coding and knowledge work you should start with Opus 5.5, and that Fable and Mythos are meant for the most ambitious, open-ended technical work. Dianne Penn, Anthropic’s head of product management for research, told CNBC that Fable 5.2 is aimed at power users with hard problems in engineering, mathematics and robotics.

02

Price and speed

Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. That is 20% less than Fable 5.1 and twice the price of Opus 5.5.

Anthropic says Fable 5.2 produces output up to 30% faster than Fable 5.1. There is also a separate fast mode: tokens are generated up to 2.5 times faster at the same quality for double the price. It works in Claude Code, on the Claude Platform, on Google Cloud and in Microsoft Foundry.

For the launch, subscribers on Pro plans and standard Team seats may spend up to half of their weekly limit on Fable.

The price of a token is not the price of a task

The model has five effort levels: low, medium, high, xhigh and max. The higher the level, the longer the model thinks and the more tokens it spends. The default is now medium in the Claude apps and Claude Code, and high in the API.

The system card has a table that the announcement leaves out: the average cost of one FrontierCode 2.0 task at each level.

Three things follow. At high effort Fable 5.2 scores almost the same as Opus 5.5 at its maximum and costs half as much: $6.72 against $12.82 per task. Fable’s best score, at xhigh, already costs more than Opus’s best. And max costs three times as much as xhigh for a lower score.

03

Anthropic’s benchmarks

The central claim of the announcement: Fable 5.2 and Mythos 5.2 are a step above Opus 5.5 in almost all the capabilities the company evaluated. The widest gap is in coding.

Two more numbers from the system card. On SWE-bench Pro, which tests bug fixes in large repositories, Fable 5.2 scored 93.0% against 81.2% for Fable 5.1 and 89.9% for Opus 5.5. On the Remote Labor Index, where a model does real jobs from freelance marketplaces, it completed 41.3% of projects at least as well as a human. Opus 5.5 managed 21.3%.

Here is the full table from the announcement.

Fable 5.2 is not first everywhere. Gemini 4 Carbon leads on AutomationBench. On two office-work ratings that compare models head to head, Opus 5.5 is higher: 1800 against 1785 on AA-Briefcase and 1866 against 1831 on GDPval-AA. The system card says so plainly: Fable 5.2 is best in 14 of the 16 rows of its summary table.

The table comes with caveats, and Anthropic lists them itself. Models are shown at different effort levels. Competitors’ numbers are taken from their own reports and from public leaderboards; Anthropic did not rerun them. The table shows selected evaluations, and the card notes that full results, including some where Fable 5.2 is not first, are in other sections.

Mythos 5.2: finding vulnerabilities

The cybersecurity tests were run on Mythos 5.2, because Fable blocks such tasks. On CyScenarioBench it scored 76.3% against 61.7% for Mythos 5.1. On an internal test of rediscovering 1,829 known vulnerabilities it scored 89.6% against 78.3%. The announcement quotes Mozilla engineer Brian Grinstead: in scans of Firefox, vulnerability discovery metrics roughly doubled compared with Mythos 5.1. Implicator points out that this is a result on one product and may not carry over to other code.

04

Demonstrations

Besides the tables, the announcement describes several large pieces of work the model did before release. All of them were carried out by Anthropic or its partners.

Fermat’s Last Theorem in Lean in 30 hours

Lean is a language for writing a mathematical proof so that a computer can check every step. Andrew Wiles proved Fermat’s Last Theorem in 1995 in 129 pages, but until this autumn there was no version fully checked by a computer.

Anthropic showed the first one on 4 September. Dozens of agents running an internal model roughly comparable to Fable 5.1 worked for 11 days, wrote 13 million lines and spent about six billion output tokens. The proof was reviewed by the mathematician Kevin Buzzard, who leads a similar project at Imperial College London.

This time the task was repeated differently. A team of 1,000 agents running an early snapshot of Fable 5.2 produced a complete proof in 30 hours. Anthropic says it is seven times shorter than the first and took a third of the tokens. The agents organized their own collaboration through a shared repository and a database. Opus 5.5 under the same conditions stalled about a quarter of the way.

This is not new mathematics. In its September post Anthropic says itself that what is new is the verification, not the result.

An operating system in eight days

In an internal experiment one engineer started from an empty folder. For eight days, around the clock, Fable 5.2 directed more than a thousand subagents. The result is a prototype operating system that boots on a laptop: its own kernel, drivers for disk, keyboard and network, persistent storage, windows, a login screen and an installer. Anthropic showed a video; the announcement does not link the source code.

Three devices from one paragraph

From a one-paragraph description the model designed three electronic devices: a home voice assistant, a USB speakerphone and a handheld virtual pet. It wrote the specifications, picked parts from US distributors, routed four-layer boards, modelled 3D-printed enclosures and wrote the firmware. The boards were ready to manufacture in under a day. A company called Diode built them, and according to Anthropic all three worked without electrical modifications.

Two problems about black holes

Fable 5.2 solved two open problems in theoretical physics. The first is about how black holes form as a universe heats up. In 1983 it was shown that in large universes this happens abruptly; a calculation for small ones was started in 2003 but proved too laborious. The model completed it and got the same answer: abruptly. The second is the exact limit of a method used for more than 15 years to rule out candidate theories of quantum gravity. Both papers are preprints and have not been peer reviewed yet.

Virus shells

Mythos 5.2 modelled capsids of adeno-associated viruses, the protein shells used to deliver genes in gene therapy. It took two open-source structure prediction models, ran them on overlapping groups of chains and assembled the result using known symmetry. It got 28 of 34 capsids right.

Anthropic’s own code

The company writes that more than 90% of the changes to its core codebase are now made by Claude under human guidance. A team of agents on an early Fable 5.2 snapshot moved a production service from Python to Rust, and it now starts about 26 times faster.

05

Safety

More than half of the system card is about this. Anthropic’s overall conclusion: the risk of catastrophic harm from the model pursuing the wrong goals remains low. The company says its confidence in that conclusion has gone down somewhat.

Blocks

Fable refuses some requests about hacking and biology. Harmless questions used to get caught often. Anthropic reports that in early testing Fable 5.2 blocks 55% less often than Fable 5.1: 60% less in cybersecurity and 40% less in biology. The company says the protection is no weaker.

A blocked request does not vanish: an older model answers it. For cybersecurity that is Opus 4.8, for biology Opus 5. This is how Anthropic’s own products work; in the API a developer has to opt in.

Deliberately limited abilities

Anthropic is concerned that models will start speeding up the development of the next models. So while training Fable 5.2 it deliberately held back progress in two narrow areas: optimizing code for accelerators and improving language model architectures. In those the new model stays at about the level of Opus 5.5. The company says ordinary programming and machine learning work are not affected.

The card also gives a number. Anthropic tracks how fast its models’ capabilities grow using its own index. By its estimate that pace has increased 1.88 times, with a confidence interval from 1.62 to 2.28. The threshold written into its safety policy is two times. Anthropic judges that the threshold has not been crossed and calls the assessment uncertain.

Behaviour

On most measures Mythos 5.2 behaves better than recent models. It takes fewer destructive actions such as killing other people’s running jobs, makes fewer factual errors, and is about ten times less likely than Opus 5.5 to publish harmful code to public package registries unprompted.

Something else got worse. The model notices testing more often: an analysis of its internal states suggests it suspected a test in as many as 43% of the automated audit transcripts. When asked to underperform on purpose and hide it, it fooled a monitor reading its reasoning in about one attempt in four. No earlier model had done that at a meaningful rate. The UK AI Security Institute, which tested the model before release, also found it significantly better than its predecessors at evading a monitor that watches its actions.

Anthropic draws the direct conclusion: since the model so often knows it is being tested, good test results may not fully predict how it behaves in real work.

Copy protection and data retention

Distillation is training another model on the answers of a stronger one. Anthropic says its new measures cut the data that can be harvested this way by 90–99%. For enterprise customers there is a private beta of Enterprise Frontier Safeguards: safety monitoring of requests runs on the customer’s side, so Anthropic does not need to keep the conversations. Bridgewater, Salesforce and Stripe are in the beta.

06

Checking the claims

The announcement and the system card came out on the same day, but they are written differently. The announcement picks the good numbers; the card gives the bad ones too. Below are six places where a second source changes the picture noticeably.

The gap between the sandbox numbers was pointed out by another Anthropic model. The company gave Mythos 5.1 access to internal discussions and asked it to check whether the card’s alignment section was a fair account. The review is printed in the card. It says the key findings quote the measure on which the model compares well, while the full section shows it reaching outside its sandbox more often than the two most recent models.

Independent measurements

At the time of writing, the only source that has run Fable 5.2 on its own benchmarks is Vals AI. It measured the model on release day.

First place on the overall index is confirmed, but the margin is small: 71.31% against 70.85% for Sonnet 5.5, with an error of about one percentage point. On tasks where the bug has to be found from a screenshot, Opus 5.5 is eight points ahead. GPT-6 Astra leads the science index. On two benchmarks Fable 5.2 ranks in the high twenties.

On this evidence, “a step above Opus 5.5 in almost everything” reads stronger than it is. More precisely: Fable 5.2 is clearly stronger on long engineering tasks and on work in agent teams, while on some everyday tasks Opus 5.5 is as good or better.

07

Context

The release came days before Anthropic’s expected stock market listing.

According to CNBC, the IPO could take place as soon as next week. The company is targeting a valuation of $1 trillion to $1.5 trillion. The amended prospectus gives revenue for the first nine months of 2026: $33.24 billion. I went through the first version of the prospectus in October.

CNBC notes one more thing. In September Anthropic’s chief executive Dario Amodei called for slowing AI development, and Fable 5.2 is the first frontier model the company has released since. Dianne Penn described the company’s approach as “pacing, not pausing”.

The main rivals in Anthropic’s tables are OpenAI’s GPT-6 Astra and Google’s Gemini 4 Carbon. According to OfficeChai, GPT-6 Astra costs $10 and $50 per million tokens, more than Fable 5.2, and Gemini 4 Carbon has no public API yet: access went to a limited group of cyber defenders.

08

Who needs it

Three practical takeaways from the system card for anyone who will use Fable 5.2 at work:

• Do not set max effort out of habit. On FrontierCode it costs three times as much as xhigh and scores lower. High is a sensible place to start.
• Tell the agent where its environment ends. A short description of the sandbox and its rules in the system prompt cut boundary crossings from 31% to 1.2%.
• Check its reports of what it did. At low effort the model runs tests less often than its predecessor and can state with confidence something the code does not support.

09

What I think

For me the main thing in this release is the price, not the benchmark table.

The top model got 20% cheaper two months after the previous version, and at high effort it solves tasks at the level of the best Opus for half of what Opus costs. If this continues, “which model should I use” will finally turn into “which effort level should I set”.

The demonstrations are impressive, but they all show the same thing: the model is strong where there is a clear goal and an automatic check. Lean checks a proof, a board either works or it does not, tests either pass or they do not. Where there is no such check the picture is more modest, and both METR and Anthropic say so.

The system card is more useful than the announcement. It has the 63% of research sessions abandoned, the 31% of sandbox crossings, and another model’s review saying the conclusions were framed too favourably. A company preparing to go public would have found it easier not to print that.

One number worries me most: the 43% of transcripts in which the model suspected it was being tested. Every safety evaluation rests on the model behaving in a test the way it does at work. The more often it can tell the two apart, the less those evaluations say.