Anthropic engineer on Claude Code effort levels: low to iterate, max for security
trq212 · x · 2026-09-26
Anthropic engineer Thariq Shihipar published a detailed explainer on how effort works in Claude Code, based on his own tests of three builds and a deep dive into Terminal-Bench 3.0.
- What effort is: an approximation of how much compute the model spends on a task. Higher effort means more independent judgment, more verification and edge-case testing — without breaking the prompt cache.
- When to go high: domains where verification matters most — hardware, code review, security. Effort max is mostly for fully hands-off runs or hunting security vulnerabilities.
- When to go low: when you want to stay in the loop.
- His workflow: have the model interview him first, implement on low effort, review the output manually, then run verification on high effort.
Related event: Anthropic Explains Claude Code Effort Setting(4 posts)→
More from coding & agent
- "Coding is solved": dev says he never reads a single line, relies on model audits — Yamapama · 2026-09-26
- mitsuhiko agrees Agent hype mirrors the OC hype cycle from eight months ago — mitsuhiko · 2026-09-26
- Kev: open-source Jev-like decision models from 0.8B to 27B run locally for free — alexcovo_eth · 2026-09-26
- Having AI build its own asset creation tools first pays off for 3D scene generation — gandamu_ml · 2026-09-26
- Tev1 0.8B open-sourced: a tiny Jev-like classifier running locally on Mac at ~50ms — iamrobotbear · 2026-09-26
- Armin Ronacher: with agents this good, Rust devs should stop fearing unsafe — mitsuhiko · 2026-09-26