Tech • AI • Robotics • Game
Anthropic's New Fable 5.2 Crushes Opus 5.5 and Ranks First Worldwide
9.4/10TL;DR
Anthropic launched Claude Fable 5.2 and the restricted Mythos 5.2, claiming major gains in long-horizon, tool-heavy reasoning, coding, robotics and scientific work as the company heads toward a closely watched IPO.
KEY POINTS
A faster formal math result
In a headline demonstration, 1,000 Fable 5.2 agents were given a minimal prompt and produced a Lean formalization of Fermat’s Last Theorem in about 30 hours. Anthropic said an earlier effort with an unreleased model, a specialized harness and researcher guidance had taken 11 days. The new run used about one-third as many action tokens and yielded a proof roughly seven times shorter.
Clear gap with Opus 5.5 on long runs
Anthropic also ran 1,000 Opus 5.5 agents on the same task. At 18 hours, Fable had reached 86% completion while Opus was at 26%; by 30 hours, Fable had finished while Opus had moved only to 27% and stalled. The result reinforced Anthropic’s broader claim that the new system’s gains are concentrated in sustained, multi-agent workflows rather than quick single-turn answers.
Two versions with different safety access
Fable 5.2 is the public model and includes cyber and biology guardrails, while Mythos 5.2 removes those restrictions for vetted security and life-science institutions through Anthropic’s verification program. Anthropic said the public model wrongly refuses requests about 55% less often than before. Fable is available immediately, including on the $20 Pro plan through November 16 with expanded reset allowances for longer-term subscribers.
Benchmarks show uneven but significant gains
On Artificial Analysis’s Intelligence Index V5, Fable 5.2 scored 65, ahead of Opus 5.5 at 59 and far above Fable 5.1 at 49. Anthropic highlighted jumps in Robo Dojo from 27.6 to 64.9, chart understanding from 44.8 to 73.4, remote professional tasks from 17.9 to 41.3, terminal-based science from 52.6 to 74.6, and Humanity’s Last Exam with tools from 65.6 to 74.2. By contrast, lighter gains appeared on OSWorld and DoubleA Briefcase, suggesting the biggest advances come when models must plan, use tools and persist over time.
Over-deliberation complicates coding economics
On Frontier Code 2.0, Fable scored 49.1% at extra-high effort for about $14.19 per task, topping GPT-6 Astra at 42.8%. But at maximum effort it fell to 45.6% while cost rose to $44.70. Anthropic described this as over-deliberation: the model can pull in too much context, over-refactor simple fixes and even think past the time limit. Independent analysis later argued that, in practice, max mode was often too verbose and more expensive than Anthropic’s framing implied.
Lower token pricing, mixed per-task costs
Anthropic priced Fable 5.2 at $8 per million input tokens and $40 per million output tokens, down from $10 and $50 on Fable 5.1, while claiming speeds up to 30% faster. A faster mode costs $16 and $80 for up to 2.5 times the speed. Anthropic said Fable at high can match Opus 5.5 at max for roughly half the per-task cost, but external testers found some long runs expensive because the model generated exceptionally large token counts.
Coding, systems and hardware demos broadened the case
Engineers reportedly gave Fable an empty folder and asked for an operating system; after eight days and more than 1,000 sub-agents, it produced a bootable system with a kernel, drivers, persistent storage, a windowed interface and login screen. Another team rewrote a production service from Python to Rust, cutting startup time by 26 times. Anthropic said more than 90% of merged code in its core codebase is now written by Claude under human supervision.
Science, biotech and robotics stood out
Anthropic said Fable solved two open physics problems tied to holographic duality, with results now in preprints. In biotech, Mythos 5.2 built its own pipeline for adeno-associated virus capsids and correctly identified 28 of 34 natural and engineered capsids in work with Dino Therapeutics. In robotics, Anthropic’s internal Robo Dojo max-reasoning run showed Fable at 64.9, versus Opus 5.5 at 39.4 and GPT-6 Astra at 29.8, while third-party demonstrations showed recovery from grip errors and changing hardware conditions during a full 62-move checkers game.
Safety limits and IPO questions remain central
Anthropic said it deliberately capped Fable’s ability to accelerate AI research, keeping some model-improvement and accelerator-kernel skills around Opus 5.5 level and adding anti-distillation measures said to cut scraping by 90% to 99%. Meanwhile, Mythos 5.2 improved CVE rediscovery by 10.2% over the previous best and reduced sandbox breakout attempts to 3.1%, down from 39% on Mythos 5.0. The launch arrived as reports around Anthropic’s IPO valuation shifted sharply, with figures ranging from roughly $867 million shares at $70 to $80 to investor expectations spanning well above $1 trillion.
CONCLUSION
Fable 5.2 appears to be a significant step forward for persistent, tool-using AI work, especially in coding, science and robotics. The immediate question is whether those gains can translate into reliable economics and governance as model capability and market pressure rise together.
Ask a question