Anthropic releases Claude Fable 5.2: 65 points in the ranking, but $26.50 per task

Anthropic has taken back first place in the rankings shortly before its IPO. Claude Fable 5.2 and Mythos 5.2 share a single base and are designed for the heaviest technical workloads: scientific computation, vulnerability discovery and long agentic chains. For everyday work, the company still recommends Opus 5.5, and that caveat matters: the new model is not built for every scenario.

Numbers and price

On the Artificial Analysis v5 index at maximum compute, Fable 5.2 scores 65 points. Gemini 4 Carbon, which Google has not yet released, scores 60, and Opus 5.5 scores 59. The previous model, Fable 5.1, released in early September, trails by 15 points. Fable 5.2 is the best in five of nine tests: 80% on HLE-Diamond (expert-level academic questions), 73% on Terminal-Bench-Science and 49% on FrontierCode 2.0, where an agent must make useful changes to real open-source projects.

The price per million tokens has dropped: $8 for input and $40 for output, down from $10 and $50. But the bill depends not on the token price but on how many tokens the model spends. At peak settings, one task costs $26.50 on average. That is 2.6 times more than Opus 5.5 ($10) and roughly double Fable 5.1 ($12.99).

The lower modes are more interesting. At the High and Xhigh levels, Fable 5.2 scores 62 points, more than any other model at any level, and spends just 98 thousand output tokens per task. Opus 5.5 at the Max level spends 173 thousand. If you need quality above Opus but have a limited budget, this is where to look. There is also a fast mode: output is up to 2.5 times faster, but twice as expensive. It is available to select organizations on Claude Code, Claude Platform, Google Cloud and Microsoft Foundry. Without it, the model generates text about 30% faster than its predecessors, according to Anthropic.

Where the lead ends and what to do about it

Fable 5.2 is not a universal winner. In Terminal-Bench 4.0, where the model performs multistep tasks in a terminal (development, security), it scores 60%. That matches Opus 5.5 but is below Kimi K3.1 Pro (67%), Gemini 4 Carbon (65%) and Sonnet 5.5 (64%). In AA-Briefcase, a benchmark of multistep office work, it scores 1785 Elo against 1813 for Sonnet 5.5. The gap to Opus 5.5 (1800) falls within the confidence interval, meaning it is statistically a tie.

The practical takeaway is simple. Don't switch all your traffic to Fable 5.2. Run your hardest tasks on it: scientific work, research, long agentic chains with real repositories. Compare not the price per token but the total cost of a solved task on your own test set. For routine terminal coding and office automation, Opus 5.5 or its competitors may be both cheaper and no worse.

Mythos, Fermat and a brake on AI self-acceleration

Fable is available to everyone, but it blocks risky cybersecurity and biology requests, such as pentests and exploit generation. Mythos lifts some of the restrictions and is open only to organizations vetted through Anthropic's cybersecurity or life sciences programs. In the internal CVE Rediscovery test, where the model searches for a bug in a project's pre-fix state across 1829 known vulnerabilities, Mythos 5.2 found 10.2% more than Opus 5.5 and 14.4% more than Mythos 5.1. The lead held at every compute level. According to Anthropic, one security firm reported that every sixth confirmed bug found by Mythos 5.2 was missed by all the other models tested.

Separately, Anthropic says it deliberately did not let the models improve skills that directly speed up AI development: optimizing compute kernels for accelerators and improving LLM architectures. Here they perform roughly on par with Opus 5.5. Agentic coding, hardware development and classical research, including interpretability, were not affected by the restrictions. According to the company, Claude already writes more than 90% of the changes that land in Anthropic's main codebase, under human supervision.

The main demonstration is in mathematics. About a thousand Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem in the Lean language in 30 hours, with the computer verifying every step. The previous research version of Claude took 11 days to do this, in a purpose-built environment and with human hints. The new agents started from a short instruction, organized the work themselves and did not use mathematical sources. Token spending on actions came to a third of the first attempt, and the proof came out seven times shorter. In the same setup with a thousand agents, Opus 5.5 reached only 26–27% of the formalization and made almost no progress after the 18th hour.

All of these figures come from Anthropic itself and Artificial Analysis. There is no independent verification yet, and an outside specialist cannot reproduce internal tests like CVE Rediscovery. Treat them as a good reason to run your own measurements, not as a ready-made solution.