Claude Opus 5 peaks at medium effort on FrontierCode, while max can hurt coding quality
量子位 · wechat · 2026-07-27
Opus 5’s best coding performance shows up at medium, not max
A test on the FrontierCode benchmark swept Claude Opus 5 across inference-effort levels from low to max and found a counterintuitive curve: medium delivered the best performance, while cranking effort all the way up did not help and could even hurt.
- The article argues that higher effort mainly means more reasoning budget, not a smarter model.
- For simple tasks like extraction, classification, and documentation, low and high produced similar quality.
- In coding, especially small bug fixes, high effort often caused the model to overdo it: refactor unrelated code, rename variables, adjust imports, and turn a small patch into a broader PR.
- Anthropic’s own guidance reportedly suggests using low/medium whenever quality does not drop, reserving higher effort for genuinely hard long-horizon tasks.
- A practical caveat: effort level is part of the cache key, so switching levels mid-session can invalidate cached context and raise total cost.
The recommended strategy is to pick one effort level per workflow and keep it stable to preserve cache hits. For many structured tasks, the best tradeoff appears to be medium, not max.
More from coding & agent
- Hands-on walkthrough shows how AI agents are automating SDLC in 2026 — Pavan_Belagatti · 2026-07-27
- Why Retrieval, Reasoning, and Execution Must Remain Separate AI Layers — Harshit-24 · 2026-07-27
- Watching 10 RPA bots run on Xiaohongshu feels like hearing a baby cry — huangyun_122 · 2026-07-27
- Go SAST triage agent cuts scanner noise, but still needs humans — coldyx · 2026-07-27
- Claude Sonnet SEO workflow hits context bloat on multi-domain SERP JSON — MaTT_fromIT · 2026-07-27
- Builder shares a 17k-memory Claude harness with Postgres retrieval and nightly judging — supermegasaurusrex · 2026-07-27