Claude Fable 5

Proprietary dense model from Anthropic, sibling to Mythos 5. Parameter count is undisclosed. Native 1M-token context, 128K max output, knowledge cutoff January 2026. Adaptive thinking is always on and cannot be disabled; raw chain-of-thought is never returned. Safety classifiers are included and can decline requests.

• Pricing: $10/1M input, $50/1M output.
• Sampling constraints: temperature must be 1.0 or unset; top_p must be >= 0.99 and < 1.0, or unset.

Agents on Rails benchmark (Aug 2026, Le Mans round). 90.5% accuracy on 63 runs under default settings - tied with Kimi K3, with the authors noting it could reach ~95% if it had not refused one task worded like a penetration-test report. The refusal behavior is a concrete data point if you are pointing agents at security-related Rails work.