Anthropic's Claude Fable 5 tops benchmarks, but best scores are restricted
Fable 5 leads SWE-Bench at 80.3%, yet its highest scores come from the banned Mythos 5 model.
Anthropic released Claude Fable 5 on June 9, 2026, calling it the most capable model it has ever made generally available—but only customers in the US can access it due to a government restriction on its Mythos-class capabilities. Benchmark comparisons against GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.8 show Fable 5 leading on tasks like agentic coding (SWE-Bench Pro 80.3%) and knowledge work (GDPval-AA ELO 1932). However, according to the article, starred rows in the benchmark table show scores from the restricted Mythos 5 model, not the deployable Fable 5; for example, ExploitBench (cyber) lists 78.0% from Mythos 5, while Fable 5 made 0% progress on offensive cyber tasks in blocking mode. Pricing for Fable 5 is $10 per million input tokens and $50 per million output tokens.
- Claude Fable 5 leads SWE-Bench Pro at 80.3% and FrontierCode at 29.3%, far ahead of GPT-5.5 and Gemini 3.1 Pro.
- Top scores on cybersecurity (ExploitBench 78%) and biology (HealthBench 66%) come from the restricted Mythos 5, not Fable 5.
- Pricing: Fable 5 at $10/$50 per million tokens; GPT-5.5 and Opus 4.8 are cheaper but trail on key coding tasks.
Why It Matters
Developers must navigate region locks, benchmark asterisks, and tiered pricing to avoid costly deployment mistakes.