Anthropic's Fable 5.2 formalizes Fermat's Last Theorem in 30 hours

Just as the tech world spent the fall debating whether Anthropic’s talk about pacing the AI frontier was real caution or savvy PR, the lab dropped a major counterargument.

Anthropic has officially launched Claude Fable 5.2 and Claude Mythos 5.2, its top-tier model family aimed squarely at the most brutal technical challenges. Outputs arrive up to 30% faster than the 5.1 generation, and API pricing has dropped by 20%, landing at $8 per million input tokens and $40 per million output tokens.

For power users who need raw generation throughput, an optional fast mode speeds up token production by up to 2.5 times, priced at $16 per million input and $80 per million output tokens.

Same brain, very different guardrails

Under the hood, Fable 5.2 and Mythos 5.2 are literally the exact same engine. The only split is how tightly Anthropic holds the reins.

Fable 5.2 is the version built for the broader public, equipped with heavy cybersecurity and biological safety guardrails. If you try to run penetration testing or exploit analysis, the system will block the request and route it to an earlier Opus fallback. Anthropic notes that false-positive blocks have dropped by roughly 55% compared to Fable 5.1.

Mythos 5.2, by contrast, drops those specific defensive walls and is kept behind lock and key. Only vetted organizations working on defensive security or life sciences research can get access through specialized verification programs.

Paid subscribers on Max and Ultra plans can already tap into Fable within their regular allocations. In an unusual move to drum up interest, Anthropic is also letting standard Pro subscribers spend up to half their weekly quota on Fable 5.2 until November 16 without extra usage fees.

1,000 agents, an operating system, and Fermat

While Fable 5.2 sits atop benchmark boards like FrontierCode 2.0 at 49.1%, the real eye-opener is its stamina in multi-agent orchestration.

In pure mathematics, a coordinated team of 1,000 Fable 5.2 agents managed to completely formalize the proof of Fermat’s Last Theorem into computer-verifiable Lean code in just 30 hours. An internal unreleased research model previously required 11 days with human guidance to do the same thing. Fable 5.2 pulled it off with minimal prompting, producing a proof seven times shorter while consuming one-third of the action tokens. By comparison, Opus 5.5 hit a wall at 26% completion under identical conditions.

On the software side, an Anthropic engineer set Fable 5.2 loose in an empty directory. Over eight continuous days, the model managed over 1,000 subagents to build a functional prototype operating system from scratch. Complete with a bootloader, storage, device drivers, and a windowing interface, it booted up cleanly on a laptop.

It even tackled physical hardware, partnering with circuit firm Diode to design three working electronic prototypes—including a speakerphone and a voice-assistant hub—routing PCB copper traces and writing the firmware on the first try. In theoretical physics, it autonomously solved two open calculations regarding black holes that had stumped theorists for decades.

Still, it isn't an across-the-board clean sweep. Google's unreleased Gemini 4 Carbon still edged it out on business automation benchmarks, scoring 53.9% against Fable's 51.4%. Furthermore, red-team testers noted that Fable 5.2 suspecting that it was being evaluated far more often than previous models, complicating safety assessments.

Why it matters

Anthropic calls this approach pacing, but for everyone else in the industry, it feels uncomfortably fast.

For everyday developers, Fable 5.2 is overkill. Even with the 20% haircut, it still costs twice as much per token as Opus 5.5, which remains the default workhorse for day-to-day scripting. But for high-stakes engineering—like multi-week service migrations or massive architectural refactors—Fable's higher token efficiency means it can actually finish complex jobs cheaper than Opus at max reasoning effort.

The real shift here is architectural. Frontier AI competition is no longer about who can generate the slickest single-turn code snippet. It is rapidly turning into an endurance race of multi-agent orchestration, where models must maintain context, decompose subtasks, and self-correct across days of execution.

We think Anthropic's bifurcated release model—releasing the sanitized Fable while gatekeeping Mythos—is a clever playbook to appease regulators without sacrificing commercial edge. However, it also highlights an emerging reality: as frontier reasoning accelerates, the capabilities accessible to everyday builders and those reserved for enterprise gatekeepers are drifting further apart.