source: simon willison: initial impressions of claude fable 5
level: technical
claude fable 5 launched alongside claude mythos 5, with anthropic claiming equal performance but stricter safety guardrails. the models offer a 1 million token context window, 128,000 max output tokens, and a knowledge cutoff of january 2026. pricing is double that of claude opus 4.8 at $10 per million input tokens and $50 per million output tokens, with no extra cost for long context. the safety classifiers in fable trigger often, and the api includes new fallback options when content is rejected.
the model feels large, showing extensive knowledge of simon willison's open source projects when prompted, far surpassing opus 4.8's response. while willison typically values tool use over baked-in knowledge, he notes that deeper world knowledge may improve coding tasks by providing better awareness of libraries and patterns. anthropic has not disclosed model size, but speed, pricing, and knowledge suggest fable may be one of the largest models available.
in practical tests, fable upgraded a micropython sandbox to run full python via webassembly, producing a working wheel file. it also implemented a human-in-the-loop approval feature for datasette agent, leading to significant improvements in the llm library, including pause-resume tool chains and better tool call handling. willison spent $110.42 in tokens during testing, all within his subscription, and noted the model's strong api design, testing, and documentation output.
why it matters: larger models with deeper knowledge may improve coding and tool-use tasks, but higher costs and strict safety filters could affect practical deployment.
source: simon willison: initial impressions of claude fable 5