Claude Fable 5 is Anthropic's first generally available Mythos-class model, and Claire Vo got early access before launch. The verdict is mixed. It crushes benchmarks including SWEBench Pro, ships alongside a new Managed Agents product, and introduces a safety fallback concept built into the architecture. It is also token-intensive by design, which means costs scale fast in real workloads.
Vo ran three practical tests: writing a product graph spec, designing a skills registry, and orchestrating a multi-agent pipeline. The model performed well on complex reasoning and specification work but showed a consistent pattern of caution at execution time. It thinks carefully, then hesitates to act. That tradeoff matters depending on where you deploy it.
The episode is worth watching in full for the safety classifier and fallback discussion at 6:28 and the multi-agent test at 14:43. Those two segments reveal how Fable 5 behaves at the boundaries that most benchmark scores ignore entirely.
[READ ORIGINAL →]