Anthropic Staff Explains How to Choose Claude Code's Effort Setting
Anthropic employee Thariq Shihipar (trq212) published a detailed post explaining how Claude Code's effort parameter works and when to use each level, drawing on his own testing of three model versions and Terminal-Bench 3.0 data. Media outlets like Every and AI寒武纪 followed up with summaries.
Confirmed
- Effort is essentially a compute-budget knob: it adjusts how much verification, edge-case testing and autonomous judgment the model performs, spending more resources on checking boundary conditions and self-verification rather than fixing flawed reasoning
- Changing effort does not break prompt cache
- Guidance: low effort suits fast iteration with human oversight; high effort suits tasks where errors are costly, such as code review; max is for security-audit work
- Every relayed his view: for tasks like an HTML sanitizer, where there are too many ways to fail, the extra checks are worth the tokens
- Terminal-Bench 3.0 data shows hardware-task pass rate rising from 34% to 75% when moving from the lowest to a high effort level (relayed by AI寒武纪 as Fable 5.1)
Unconfirmed
- The model version cited for Terminal-Bench 3.0 differs across relay posts; trq212's original text is authoritative
Why it matters
- Many users don't know which effort level to pick; this hands-on guidance from a frontline employee offers a concrete selection guide directly useful to Claude Code developers
2026-09-25 ~ 2026-09-26 · 5 related posts
Primary sources
- Anthropic engineer: crank Claude's effort setting for risky code, dial it down for routine tasks — every · 2026-09-25
- Anthropic deep dive: effort tuning in Claude Code pays off most for security and code review — trq212 · 2026-09-26
- [source] Anthropic engineer on Claude Code effort levels: low for staying in the loop, max for security audits — trq212 · 2026-09-26
- Claude Code Team Explains Effort: Hardware Task Pass Rates Jump from 34% to 75% — AI寒武纪 · 2026-09-26
1 near-duplicate retellings: trq212