Claude Fable 5.1: A Million Tokens, but Check the Bill First
If your retrieval pipeline keeps dropping the one clause that matters, a 1,000,000-token window changes the design question. Anthropic: Claude Fable 5.1 lets you skip chunking for whole-repo audits, multi-contract review, or a quarter of support transcripts, and ask cross-document questions without a vector store guessing what's relevant. That is the use case where it fits: tasks where losing context costs more than a slow, heavy request.
The trade-off is that our catalog has no input or output pricing and no strength tags listed for it yet, so budget is an open question. Long-context calls multiply whatever the per-token rate turns out to be, and latency grows with prompt size. You are also tying the pipeline to one vendor's API and caching behavior. For short chat, classification, or high-volume extraction, a smaller model with known pricing is the safer default. Pick Fable 5.1 for full-corpus reasoning, after confirming its rates.