Claude Fable 5.1 ships to Claude Code and the API, cache 75% cheaper
The point release more than doubles agentic-science scores and cuts heavy agent workloads up to 45% via cheaper cache reads. Mythos 5.1 stays gated.
Copy markdown
Agentic-science scores more than doubled
Fable 5.1 posts 52.6% on Terminal-Bench-Science 0.1, up from 24.7% for Fable 5 and past Opus 5's 29.0%, alongside 55.8% on Terminal-Bench 4.0 and 73.4% on CursorBench 3.2.0. The jump shows up in long, multi-step coding and research runs, not one-shot prompts.
Cache reads cut 75%, so agent loops cost less
Cache reads drop to $0.25/M from $1/M while input ($10/M) and output ($50/M) stay flat. Anthropic pegs the net effect at around 25% off typical workloads and up to ~45% off heavily agentic ones — the loops that re-read the same context every turn.
How to switch today
It's live now as claude-fable-5-1 on the Claude API, Claude Code, AWS Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, with a 1M-token context and 128K max output. Same tokenizer as Fable 5, so migrating is mostly swapping the model string.
Mythos 5.1 is the gated, slightly stronger twin
Same architecture with lighter safeguards, Mythos 5.1 edges Fable on the hardest evals (60.9% vs 55.8% on Terminal-Bench 4.0) but is restricted to vetted US organizations in Project Glasswing. Most builders won't get keys — Fable is the one you'll ship on.