Claude Fable 5.1: A Hedge Fund's 5-Year-Old Bug, Solved
Hedge Fund Millennium: This AI Found a Crash Nobody Could Explain for Four or Five Years
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. They’re the same underlying model, differing only in safeguard level: Fable 5.1 is available to everyone, while Mythos 5.1 is restricted to vetted professional organizations. The most striking line in the announcement came from hedge fund Millennium, which said the model found “a crash that nobody on our team had explained in four to five years.”
Science-task accuracy more than doubled
The benchmarks back up that claim. On Terminal-Bench-Science 0.1, a test that simulates real scientific research workflows, the new model scored 52.6%, versus 24.7% for the previous Fable 5 — more than double. On coding, Mythos 5.1 hit 60.9% on Terminal-Bench 4.0, up from 42.0%. The knowledge-work benchmark GDPval-AA v2 rose from 1723 to 1853. Multidisciplinary reasoning with tools on Humanity’s Last Exam improved far more modestly, from 63.8% to 65.0% — not every metric doubled, which at least suggests this wasn’t a launch built around cherry-picking the flashiest numbers.
What other early users are saying
Quant trading firm Jane Street said Fable 5.1 “solves more of our coding problems than Fable 5 or Opus 5, achieving state of the art on trading intuition.” AI-engineer startup Cognition, maker of Devin, put it bluntly: “Moving Opus 5 traffic to Claude Fable 5.1 on launch day.” Database company MongoDB said its team “built a complex prototype in about three days” with “strong verification loops.” Contract-review tool Crosby’s score improved from 47.9 to 57.0, and browser-automation tasks hit an 82% completion rate.
Heavier use, faster inference, same sticker price
Anthropic didn’t raise base pricing this time: input tokens are still $10 per million, output $50 per million. What actually changed is cache-read pricing — cut from $1.00 to $0.25 per million tokens, a 75% reduction. Anthropic estimates this saves roughly 25% on typical workloads, and up to 45% on highly agentic ones.
The Mythos side: mapping Venus, designing proteins
Mythos 5.1, the limited-access version, shows off harder-core research capability. In molecular design tests, it produced molecules with binding affinities 10 times higher than competition entries, hitting a 50% success rate across 12 protein targets versus a typical 10-15%. It also took 30-year-old NASA data and recomputed a high-resolution elevation map of Venus, improving resolution from 10-20km down to 2-3km and boosting height accuracy by 25%. In computational biology, it optimized seven deep-learning models by up to 2.5x, cutting GPU costs on genome-wide analyses by 30-60%.
Safeguards got an upgrade too
As capability climbed, Anthropic also strengthened its safety layer: false positives on cybersecurity tasks dropped 60%, and the new model can now be used defensively — actively finding software vulnerabilities rather than being used to exploit them. Biology-related safeguards were also tuned to be more precise, cutting down on benign queries getting blocked unnecessarily. An enterprise offering called “Enterprise Frontier Safeguards,” rolling out fall 2026, will let customer data stay entirely on their own cloud infrastructure. For the two highest-risk domains — cybersecurity and life sciences — Anthropic set up separate verification programs; Mythos 5.1 access is currently limited to US organizations and requires coordination with government agencies. A new anti-distillation mechanism also means new users can no longer manually edit prior context while still preserving Claude’s full thinking transcripts — a technical barrier specifically aimed at training other models off Claude’s reasoning process. These models have also signed the EU AI Act’s Code of Practice (July 2026), carry an invisible watermark on their outputs, and offer a detection API to regulators, law enforcement, media outlets, fact-checkers, and researchers.
Where to get it
Fable 5.1 is already live on Claude.ai, Claude Code, Claude Enterprise, and the Claude Platform, plus AWS, Google Cloud, and Microsoft Azure — developers can call it directly via the API model ID claude-fable-5-1. Mythos 5.1, given its higher safety-review bar, is currently limited to US organizations and requires coordinating with government agencies to get access.
