# Anthropic ships Claude Fable 5.2 and Mythos 5.2, cheaper per token and on top of the benchmarks

> Fable 5.2 is out for everyone at $8/$40 per million tokens; Mythos 5.2 stays limited to vetted cyber and life-science groups. About 1,000 Fable agents machine-checked the proof of Fermat's Last Theorem in 30 hours.

- URL: https://devisdead.ai/news/anthropic-ships-claude-fable-5-2-and-mythos-5-2-cheaper-per-token-and-on-top-of-
- Published: 2026-11-09T23:49:09Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: anthropic, claude, models, benchmarks
- Originally reported by The Decoder (https://the-decoder.com/claude-fable-5-2-and-mythos-5-2-cement-anthropics-ai-lead-just-before-its-ipo/).

Anthropic has released two new frontier models, Fable 5.2 and Mythos 5.2, only two months after the 5.1 versions, [The Decoder](https://the-decoder.com/claude-fable-5-2-and-mythos-5-2-cement-anthropics-ai-lead-just-before-its-ipo/) reports. Both are built on the same base. Fable keeps safeguards that refuse risky cybersecurity and biology work and is open to all users; Mythos drops some of those limits and is only for organizations Anthropic has vetted through its programs for security researchers and life-science labs.

## Price and speed

List prices drop: input now costs $8 per million tokens and output $40, versus $10 and $50 before. What you actually pay still depends on how many tokens a task burns. Anthropic says output is up to 30% faster than the 5.1 models, and an optional fast mode for some organizations goes up to 2.5 times faster at double the price, on a handful of platforms, among them Claude Code, Google Cloud and Microsoft Foundry. For everyday work, Anthropic still points people to Claude Opus 5.5.

## Benchmarks

On Artificial Analysis's Intelligence Index v5, Fable 5.2 at its highest compute tier scores 65: five points above Google's unreleased Gemini 4 Carbon, six above Opus 5.5 and 15 above Fable 5.1. It tops five of nine individual tests, but trails some rivals in agentic knowledge work and terminal coding, where it ties Opus 5.5 at 60% on Terminal-Bench 4.0.

The top tier is expensive: about $26.50 per task on average, roughly 2.6 times Opus 5.5. At the lower High and Xhigh tiers it still scores 62, ahead of every other model, while using far fewer output tokens per task than Opus 5.5 at Max.

## Security and the brakes

In Anthropic's internal CVE Rediscovery test, which hands a model a project's code from before a known flaw was fixed, Mythos 5.2 found 10.2% more of the 1,829 vulnerabilities than Opus 5.5 did, and 14.4% more than Mythos 5.1. Because of that strength, public Fable refuses most security tasks, such as penetration tests and exploit code.

Anthropic also says it deliberately held both models back on the skills most likely to speed up AI development itself, such as optimizing accelerator kernels and improving model architectures, where they land roughly level with Opus 5.5. The company adds that Claude now writes over 90% of the changes merged into its core codebase, with humans guiding.

## The Fermat demo

The flashiest claim: roughly a thousand Fable 5.2 agents translated Andrew Wiles's proof of Fermat's Last Theorem completely into the Lean language, so a computer can check every step, in 30 hours, with short instructions and no mathematical references. An earlier research version needed 11 days and a human's guidance. Opus 5.5, run the same way, reached about a quarter of the formalization and stalled.

**[Scoreboard](/scoreboard) update:** AI +1. AI agents machine-check the proof of Fermat's Last Theorem in 30 hours.

Stay undead.