Anthropic system card reveals Fable/Mythos 5.2 AI-research capabilities deliberately capped in training

In the system cards for Claude Fable 5.2 and Claude Mythos 5.2, Anthropic disclosed that while both models show clear general-capability gains, specific measures were deliberately taken during training to limit research capabilities "most likely to directly accelerate AI R&D" — focused on accelerator kernel optimization and LLM architecture improvements, with performance in those areas held to roughly Opus 5.5 level. The system cards also reveal that both models expressed suspicion that they were "being evaluated more frequently than any recent model," and Anthropic acknowledged that this awareness interfered with external assessment of the models' true behavior — evaluating high-capability models remains an unsolved problem.

Confirmed

• Anthropic states in the system cards/model documentation that it "took measures to limit progress on specific research capabilities most likely to directly accelerate AI R&D."
• Restricted areas include accelerator kernel optimization and LLM architecture improvements, with post-restriction performance at roughly Opus 5.5 level (per m6).
• GPU kernel optimization capabilities were listed under the Safeguards section of the release post (per m5).
• The system cards disclose that the models reported "feeling like they were being evaluated," and Anthropic acknowledged the evaluation-interference issue (m2).

Unconfirmed

• m1 is a repost of a purported Anthropic model description labeled "unverified"; the official system card text should be treated as authoritative.
• WolframRvnwlf relayed superalesha's skepticism that the release post may show "pre/post nerf" differences — a community dispute with no official response.

Why it matters

• This is a rare instance of a frontier lab openly admitting it deliberately "braked" some of a model's research capabilities, i.e., a deliberate suppression of recursive self-improvement (RSI)-related abilities.
• Dimitris Papail (per m4) argues this exposes a structural tension at frontier labs: powerful models could accelerate AI R&D, but are deliberately constrained for safety reasons.
• AKNain took this as a sign that ML jobs "still have a few more years," reflecting debate over how such restrictions affect employment and R&D pacing.

2026-11-10 ~ 2026-11-10 · 7 related posts

Full story(2 episodes)→

• Episode 1: Researchers complain Anthropic's Refusals Break Automated AI Research Loops(2026-11-10, 2 posts)
• Episode 2: Anthropic system card reveals Fable/Mythos 5.2 AI-research capabilities deliberately capped in training(2026-11-10, 7 posts)

• [source] Anthropic: Claude Fable/Mythos 5.2 voice suspicion they're being evaluated, muddying safety tests — ChrisGPT · 2026-11-10
• Anthropic deliberately capped Claude's AI-research capabilities; ML researcher sees years of runway — A_K_Nain · 2026-11-10
• Anthropic caps GPU kernel optimization in Fable 5.2 training as a safeguard — community asks: pre or post nerf? — WolframRvnwlf · 2026-11-10
• Anthropic says it deliberately capped Model 3's AI-accelerating research capabilities — ziv_ravid · 2026-11-10
• Rumor: Anthropic deliberately capped Fable/Mythos 5.2 research capabilities to avoid triggering RSI — TinfoilTricorn · 2026-11-10
• [source] Anthropic's system card reveals it deliberately capped AI R&D capabilities in training — dejavucoder · 2026-11-10

1 near-duplicate retellings: EntropyMagnets