Model capability, evaluation integrity, and what inference actually costs.
When Claude Fable 5 or Opus 5 declines a request, the API returns HTTP 200, an empty content array, and a stop_reason of "refusal". Every client that indexes content[0] without checking crashes on exactly the responses it most needs to handle gracefully.
Prompts between 512 and 1,023 tokens — written off as uncacheable for two model generations — now cache on Claude Opus 5, Fable 5 and Mythos 5. At a tenth of the input price per read, the smallest documentation change in the Claude 5 launch may be the one worth re-benchmarking first.
Fable 5 requires 30-day data retention — zero-data-retention orgs get a 400 on every request. The workaround, the scope, and the procurement consequence.
Claude 5 models think by default; the reasoning bills as output and counts against max_tokens. What flips when you swap the model string.
