Anthropic's Claude Fable 5.2 takes the top benchmark spot, at a steep price per task

Two models, one engine, different doors

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, which it aims at the hardest technical work. As the-decoder.com reported, the two share the same underlying model. What separates them is who may use them and what they are allowed to do.

Fable is open to everyone, but it has safeguards that refuse risky requests in cybersecurity and biology. Mythos relaxes some of those limits, and Anthropic offers it only to organizations accepted through its verification programs for cybersecurity or life sciences.

For everyday jobs, Anthropic still points people to Claude Opus 5.5. The new pair is the premium option for heavy lifting.

What the benchmarks say, and what they cost

Artificial Analysis, an outside firm that scores AI models, gives Fable 5.2 65 points on its Intelligence Index v5 at the highest compute setting. That is five more than Gemini 4 Carbon, which Google has not yet released, and six more than Opus 5.5. It is also 15 more than Fable 5.1, which launched only in early September.

Fable 5.2 had the best score in five of the nine individual tests. Examples include 80 percent on HLE-Diamond, a set of expert-level academic questions, and 49 percent on FrontierCode 2.0, which checks whether an AI agent can make useful code changes in real open-source projects.

It is not first everywhere. On AA-Briefcase, a test of multi-step office-style tasks, it trails Sonnet 5.5 slightly. On Terminal-Bench 4.0, which covers multi-step work in a command-line environment, it ties Opus 5.5 at 60 percent. Kimi K3.1 Pro, Gemini 4 Carbon and Sonnet 5.5 all score higher there.

The top setting is expensive. A task costs $26.50 on average, about 2.6 times Opus 5.5 and roughly double Fable 5.1. A cheaper setting still scores 62 points, which Artificial Analysis says beats every rival model at any of its settings. That setting also uses far less text: 98,000 output tokens per task, against 173,000 for Opus 5.5 at its maximum setting. A token is a small chunk of text, and models are billed by the chunk.

Pricing, speed and the security angle

List prices are $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the previous models. Your real bill depends on how many tokens a job burns, which varies by task and setup. Anthropic says output is up to 30 percent faster. A fast mode for select organizations is up to 2.5 times quicker at double the price, on some platforms only.

Security is where Mythos stands out. Anthropic says it found more software flaws than earlier models and was better at reproducing and repairing them. In an internal test using 1,829 published vulnerabilities from the CVE catalog, a public list of known software flaws, Mythos 5.2 found 10.2 percent more than Opus 5.5 and 14.4 percent more than Mythos 5.1. One security firm told Anthropic that one in six confirmed bugs it found had been missed by every other leading model it tried.

Because the public Fable model blocks most security work, including penetration tests and exploit code, the strongest security results sit behind the application process. Anthropic has not said whether access will widen.

A deliberate brake, and a proof that went viral

Anthropic says it held back, on purpose, the skills that would most directly help AI speed up its own development, a worry often labeled RSI. On tasks such as tuning low-level code for AI chips and improving language-model designs, it says the new models are about level with Opus 5.5. It says related abilities, including coding agents and interpretability research, were not affected. That claim comes from Anthropic and is not independently checked.

The showpiece was mathematics. Anthropic says about 1,000 Fable 5.2 agents turned the proof of Fermat's Last Theorem into Lean, a language that lets a computer check every logical step. Andrew Wiles presented a proof of the theorem in 1995, and this is the first time it has been machine-verified in full. The run took 30 hours.

An earlier, unpublished Claude version needed 11 days, a purpose-built setup and guidance from a human researcher. The new agents began with short instructions and organized themselves. Anthropic says they used a third of the action tokens and produced a proof seven times shorter. In the same setup, Opus 5.5 completed only 26 percent in 18 hours and made little progress after that.

Anthropic also says Claude now writes over 90 percent of the changes merged into its main codebase, under human guidance.