Claude Fable 5, Anthropic's new Mythos-class model, is now available on Vercel's AI Gateway via the model ID anthropic/claude-fable-5. It outperforms prior Claude models on long-running, multi-step tasks, sustains output across multi-day runs, and dispatches parallel sub-agents reliably. First-shot correctness on complex problems is meaningfully higher, and lower effort settings now match what previous Claude models produced at maximum effort.
Anthropic has shipped Fable 5 with hard blocking classifiers on offensive cybersecurity, biology, and summarized-thinking extraction, citing real misuse risk in those areas. Zero Data Retention is not supported because cumulative request patterns are required to detect certain abuse. Prompts and completions are retained for 30 days and are not used to train the model.
The full changelog is worth reading for the specifics on how Fable 5 handles parallel sub-agent dispatch and where code review and repository investigation improvements are most pronounced. Vercel's AI Gateway routes to it with no markup on provider pricing, supports Bring Your Own Key, and includes dynamic provider sorting by latency and cost.
[READ ORIGINAL →]