Anthropic Releases Claude Cost-Saving Guide: Start with lower-capacity model, upgrade to higher tier if tasks fail, slash costs by 41% while boosting accuracy

Beating AI News Flash: Anthropic has released a Claude model selection guide, advising developers not to default to the most powerful Fable 5.2. For most programming and knowledge work, start with Opus 5.5’s Medium inference strength. For clearly defined feature development and daily automation, use the more affordable Sonnet 5.5. Simple tasks like classification and information extraction can be assigned to Haiku 5.5. Large-scale code migrations, complex research, or tasks Opus cannot complete are the only scenarios where Fable 5.2 should be considered. Anthropic notes an effective cost-cutting method is letting AI audit its own results: for example, run automated tests when coding to catch and fix errors instead of using the highest inference strength from the start. In tests on the same subset of SWE-bench Pro programming benchmarks, Anthropic had Opus 5.5 first execute tasks with Low inference, then only retry failed tasks with High. The result was a ~97% pass rate at an average cost of $0.17 per task. By comparison, running all tasks directly with High inference yielded a 95.3% pass rate at $0.29 per task — the former approach reduced costs by ~41% while delivering a slightly higher pass rate. Increasing inference strength in isolation has a cost-performance ceiling: when Opus 5.5 was upgraded from High to Xhigh, its SWE-bench Pro score improved by only ~1.4 percentage points, while its cost rose 2.5 times. Anthropic advises that if a model at High cannot solve a hard problem, try a stronger model at a lower inference strength. For instance, in the FrontierCode 2.0 benchmark, Fable 5.2 at High scored 45.3%, nearly matching Opus 5.5’s 45.5% at Max, while its average cost per task was only half of Opus. Even though Fable’s per-token API price is twice Opus’s, total costs for such hard tasks may still be lower.