Claude Fable 5.2 Proves Fermat’s Last Theorem in 30 Hours
While humanity spent centuries scratching its head over legendary mathematical enigmas, an autonomous digital collective just formalized the toughest logical proofs over a single weekend. AI swarms are officially taking over foundational science.
During an experiment by Anthropic, a swarm of one thousand Claude Fable 5.2 agents completely formalized Andrew Wiles and Richard Taylor’s proof of Fermat's Last Theorem inside the Lean environment. Working strictly from initial high-level prompts and an outline of lemma dependencies, the agents distributed the sub-proofs among themselves without consulting external textbooks, cutting token expenditure threefold compared to earlier attempts and generating a formal codebase seven times more concise.
Meanwhile, heavyweight model Opus 5.5 stalled out early under identical constraints, proving only 26% of the required assertions in its first 18 hours before grinding to a complete halt. Raw computational muscle evidently folded against superior swarm coordination.
The collective intelligence test extended far beyond abstract lemma-checking. A dedicated squad of early Fable 5.2 instances ported an entire critical production service from Python to Rust, achieving a 26-fold reduction in service boot times. Following that, a separate swarm of one thousand sub-agents started with an empty directory and spent eight days engineering a functional operating system complete with disk drivers, network stacks, a window manager, and an installer that successfully booted on bare-metal laptop hardware.
On the industry benchmark FrontierCode 2.0, Fable 5.2 reached a 49.1% success rate on the highest effort tier, matching the peak output of Opus 5.5 at roughly half the API expense. Its restricted security sibling, Mythos 5.2, simultaneously hunted down 89.6% of known software vulnerabilities in a retro-test of 1,829 historical security reports.
Both models landed across AWS, Google Cloud, and Microsoft Azure at a baseline price of $8 per million input tokens and $40 per million output tokens, alongside a high-velocity mode generating tokens 2.5 times faster.
Complex system engineering and theoretical mathematics have abruptly degraded from grueling lifetime human pursuits into routine weekend cron jobs for token mills. The bar for cognitive labor has shifted into an orbit where biological brains without automated swarms simply cannot keep up.
Source: Anthropic
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.