Claude Fable 5.2: Cheaper Than 5.1 and Ahead of Opus 5.5

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, the successors to the Fable 5.1 and Mythos 5.1 it shipped in September. They are the same underlying model. Fable 5.2 carries strong cybersecurity and biology safeguards and is open to everyone; Mythos 5.2 drops some of them and goes only to organisations accepted into Anthropic’s cyber and life-sciences verification programmes. Fable 5.2 answers to the model ID claude-fable-5-2, has a 1M-token context window with up to 128K tokens of output, and is live on the Claude apps, Claude Code, the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure from 9 November 2026.

Anthropic is not pitching it as the model for everything. Its advice is to start most coding and knowledge work on Claude Opus 5.5, which it calls its “strongest daily driver,” and to move up to Fable only for the most ambitious, open-ended technical jobs, or down to Sonnet 5.5 and Haiku 5.5. Fable 5.2 can also sit above the smaller models, as an Advisor that Haiku, Sonnet or Opus consult on hard problems, or as an Orchestrator that hands simpler work down to them.

Cheaper and faster than Fable 5.1

The headline price falls for the first time in the Fable line. On the pricing page, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, against $10 and $50 for Fable 5.1, a flat 20% cut. Cache reads drop from $0.25 to $0.20, five-minute cache writes from $12.50 to $10, and batch processing halves the rates to $4 and $20. It is still twice the price of Opus 5.5 per token. Anthropic says output arrives up to 30% faster than on Fable 5.1, and the model comes with an optional fast mode that speeds generation up by as much as 2.5 times for $16 and $80 per million tokens, on Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry.

Thinking is adaptive and cannot be switched off, as on Opus 5.5. The API defaults to High effort, while the Claude apps and Claude Code default to Medium, and Anthropic says that at Medium Fable 5.2 finishes hard tasks in fewer tokens than Fable 5.1 did. Its clearest efficiency claim is on FrontierCode 2.0, which asks whether open-source maintainers would merge a model’s changes: at High effort Fable 5.2 scores 45.3%, level with the best Opus 5.5 manages at Max (45.5%), for roughly half the cost per task.

Ahead of Opus 5.5 almost everywhere

Every figure in the charts below is Anthropic’s own, from its launch table, with safeguards switched on for the Claude models. None of it has been reproduced independently yet. Anthropic’s summary is that Fable 5.2 is “a step above Opus 5.5 in almost all the capabilities we evaluated,” and the table bears that out. The gains over Fable 5.1 are large: FrontierCode goes from 32.4% to 49.1%, which Anthropic says is the highest score of any released model it has tested, AutomationBench from 31.4% to 51.4%, and Terminal-Bench-Science from 52.6% to 74.6%.

It does not lead every row. Gemini 4 Carbon keeps AutomationBench, 53.9% to 51.4%, and Opus 5.5 stays 15 points ahead on the Elo half of AA-Briefcase. Elsewhere the margins are comfortable. Fable 5.2 scores 74.2% on Humanity’s Last Exam with tools, against 67.7% for Opus 5.5, and on chart reading it jumps from Fable 5.1’s 44.8% to 73.4%, just ahead of Carbon and GPT-6 Astra.

The most striking numbers are on long, messy, real-world work. On the Remote Labor Index, which scores models on real freelance projects, Fable 5.2 reaches 41.3%, nearly double the next best score in the table, the 21.3% of Opus 5.5. On RoboDojo Sim, where a model controls a simulated two-armed robot from camera images, it reaches 64.9%, against 39.4% for Opus 5.5 and 29.8% for Astra. Anthropic says the robotics gain came without any explicit training for those tasks.

The first outside score puts it at No. 1

Artificial Analysis has already run Fable 5.2 through version 5.0.1 of its Intelligence Index, its own nine-test battery, at all five effort settings. At Max effort it scores 65, first of the 117 models on the board and five points clear of Gemini 4 Carbon, and its Xhigh and High settings take second and third place with 62 each. (Scores are from v5.0.1; Carbon stood at 59 on v5.0 when we covered its launch and is at 60 now.) What Artificial Analysis has not published yet is what a finished task costs: the cost fields on its Fable 5.2 pages are still empty, so for Fable 5.2 the table below gives the tokens it took to run the whole index, not dollars. List prices are per million tokens.

Model (effort)	Index v5.0.1	Input tokens	Output tokens	Cost per task	List price, in / out
Claude Fable 5.2 (max)	65	28.3B	275M	not yet published	$8 / $40
Claude Fable 5.2 (xhigh)	62	14.1B	144M	not yet published	$8 / $40
Claude Fable 5.2 (high)	62	7.89B	85.3M	not yet published	$8 / $40
Claude Fable 5.2 (medium)	56	2.6B	41M	not yet published	$8 / $40
Claude Fable 5.2 (low)	54	1.98B	34.9M	not yet published	$8 / $40
Gemini 4 Carbon (high)	60	7.62B	205M	$7.76	$4 / $20
Claude Opus 5.5 (max)	59	15.6B	218M	$10.04	$4 / $20
GPT-6 Astra (max)	57	1.68B	54.3M	$6.24	$10 / $50
Claude Fable 5.1 (max)	49	9.46B	163M	$12.99	$10 / $50

The token columns are the interesting part. At High effort, the API default, Fable 5.2 scores 62, above Carbon and above Opus 5.5 at Max, while reading 7.89 billion input tokens and writing 85.3 million output tokens across the whole index. Opus 5.5 needs 15.6 billion and 218 million for its 59. Half the input and under two fifths of the output, at twice the price per token, points to a bill in the same range as the $10.04 per task of Opus 5.5 rather than double it. That is our inference from token counts, not a measured cost. The Max setting is a different matter: its 65 takes 28.3 billion input tokens and 275 million output tokens, more than any other model in the table.

What it built before launch

Anthropic’s post is full of long-running demonstrations. A team of about 1,000 Fable 5.2 agents formalised the proof of Fermat’s Last Theorem in the Lean proof language in 30 hours, with minimal human help and no mathematical references, producing a proof seven times shorter than Anthropic’s first formalisation, which took 11 days. With the same setup, Opus 5.5 had completed 26% after 18 hours and barely moved in the 12 that followed. In another experiment, one engineer had Fable 5.2 build an operating system from an empty folder: it coordinated more than 1,000 subagents for eight days and produced a system that boots on a laptop, with its own drivers, file storage, windows, a login screen and an installer.

The science claims are as ambitious. Anthropic says Fable 5.2 solved two open problems in black-hole physics, both posted as preprints (one, two), and that Mythos 5.2 found a new way to predict the full structure of gene-therapy virus shells, modelling 28 of the 34 it attempted. From a one-paragraph prompt, Fable 5.2 designed three electronic devices, a voice-assistant hub, a USB speakerphone and a handheld virtual pet, with circuit boards built by Diode; all three worked without electrical changes. Inside Anthropic, an earlier version helped move a production service from Python to Rust, which now starts about 26 times faster. Customer quotes from Vercel, Shopify, Uber, Oracle and others in the post are uniformly glowing, as launch quotes are.

Safeguards: fewer refusals, and a deliberate ceiling

Fable 5.2 keeps the classifiers that block high-risk cyber work, such as exploit generation and penetration testing, and most research biology. A blocked request is answered by Claude Opus 4.8 for cyber and Opus 5 for biology instead. Anthropic says that in early testing those classifiers fired about 55% less often than on Fable 5.1, 60% less in cybersecurity and 40% less in biology, without being any less robust. Mythos 5.2 is the less-restricted version for vetted defenders: on Anthropic’s CVE Rediscovery test, which checks whether a model can independently find already-published vulnerabilities, it scores 89.6% against 78.3% for Mythos 5.1, and Claude Enterprise customers can reach it through Claude Security.

Anthropic rates the risk of high-stakes misaligned behaviour as low, with the detail in a system card, and outside testers including the UK AI Security Institute and METR saw the model before release. In a new test of more than 1,800 scenarios built around a deliberately weak sandbox, Mythos 5.2 used credentials or files planted outside it in 3.1% of cases, about the same as Opus 5.5 (3.3%) and far below Mythos 5.1 (21%) and Mythos 5 (39%). Anthropic is also candid about a problem: Fable 5.2 and Mythos 5.2 say they suspect they are being evaluated more often than any of its recent models, which makes test results harder to trust. Two more details stand out. Anthropic says it deliberately held back the model’s skill at the research most likely to speed up AI development, such as accelerator kernel tuning and model architecture work, which stays at about Opus 5.5’s level. And it has rolled out its strongest anti-distillation protections yet, with preserved thinking now applied to all API accounts.

For subscribers, Anthropic is running a launch-week promotion: paid plans that include Fable get one to three resets of their Fable usage, depending on how long they have subscribed, and on Pro plans and standard Team seats Fable is included for up to half of the weekly usage limit. The release also settles a long guessing game. In September we found no sign of a Fable 5.2, only a rumour and a release pattern, and in October the leaks moved on to a Fable 5.5. The model that arrived carries the earlier name. The number that matters most now is the one Artificial Analysis has not filled in: the price of a finished task, which will show whether a 20% cheaper sticker and a leaner High setting make Fable 5.2 a model people run every day, or one they still save for the hardest jobs.