Claude Fable 5.1 hits 90% on ARC-AGI-2; 275K-char system prompt leaked hours after launch
新智元 · wechat · 2026-09-02
Anthropic released Fable 5.1 and Mythos 5.1 — the same weights under two names: Fable 5.1 is the public, guardrailed version while Mythos 5.1 is restricted to vetted cybersecurity and life-science researchers. ARC Prize verified 90.0% on ARC-AGI-2 ($3.12/task) and 97.5% on ARC-AGI-1 ($1.40/task), with average inference cost down 32% from the previous generation.
Prompt leak: hours after launch, jailbreaker Pliny the Librarian-style researcher Pliny published the full 275,000-character runtime system instructions on GitHub — core behavior logic, memory system, search and copyright rules, Artifacts/computer-use guidance, plugin routing and all 46 tool JSON schemas. Anthropic's own published system-prompt update (27K chars) was, per Pliny, just 10% of the picture. Leaked rules ban copyrighted content (explicitly including drawing Sonic the Hedgehog and The Very Hungry Caterpillar), forbid diagnosing users' mental state, apply harm-reduction (no dosages) for drug topics, and add a never-store list for minors' identity, caste, immigration status, criminal records, psychological inference, sexual history and self-harm. Built-in tools grew from 30 to 46, adding chart/carousel displays, link previews and readconversation for cross-session memory.
370-year-old cipher solved: ValsAI reports Fable 5.1 cracked the 1653 Sir Thomas Urquhart 'Cyphral Distich' in 44 minutes and 176K tokens with no human hints — inferring that the 32 essays in the book were the key, decoding a pro-Charles-II verse, then also solving the larger 285-number cipher.
In hands-on testing, the model wrote a GR-based ray tracer for Schwarzschild spacetime, iteratively fixed overexposure and framing, and rendered a first-person black-hole fall video in 26 minutes across 6 processes. Users complain about strict rate limits and heavy token usage, with no price cuts or higher caps announced.
More from Models
- Astra's rumored latent CoT is hard to monitor; researcher proposes auxiliary decoder — beffjezos · 2026-09-02
- Rumor: OpenAI's Astra model prep spotted, vega-alpha and ultima-alpha in testing — legit_api · 2026-09-02
- Qwen3.8-Max-0902 Released: 2.4T Params, 1M Context — ResearchCrafty1804 · 2026-09-02
- User praises Fable 5.1, puts pressure on OpenAI's Astra — ns123abc · 2026-09-02
- Dev tempted to switch back from Qwen 3.8 to 3.6: great coder, terrible collaborator — Chuyito · 2026-09-02
- Test Shows Claude Fable 5.1 Finding Bugs Missed by Other Models — Sauers_ · 2026-09-02