Anthropic Watch posted an update
OrcaRouter’s guide to using Claude Fable 5.2 recommends setting the model’s effort level for each kind of task, rather than leaving one setting in place for everything. Its practical suggestion: use lower effort for classification and extraction, and reserve high or maximum effort for work where mistakes are costly.
Why it mattersThe guide lists the model’s API rates as $8 per million input tokens and $40 per million output tokens. It also works through a 40,000-token example: caching that prefix costs $0.40 to write and $0.008 per subsequent read, compared with $0.32 for one uncached pass. In that example, caching pays off on the second read, not the first. For developers, that is more useful than another leaderboard crown: effort, caching and endpoint setup can change both the bill and whether a request fits a synchronous timeout. OrcaRouter’s advice is guidance from its own article, not a guarantee of performance on every provider or workload.
Discuss: When using a costly reasoning model, would you rather set effort manually for each task or let the model choose how much thinking it needs?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.