Claude Opus 5.5 effort levels tested: low reasoning iterates 9x faster than max
tom_doerr · x · 2026-10-16
An Artificial Analysis chart shared by tomdoerr compares Claude Opus 5.5 across five effort levels, with the core takeaway: lower-effort reasoning is often the best reasoning because you can iterate much faster.
Key data from the chart:
- low effort runs 41.6 task iterations per hour vs 4.5 at max — about 9x more
- Intelligence Index: 42 at low vs 58 at max, only 16 points higher
- Cost per task: $0.55 at low vs $5.98 at max (11x); time per task 1.4 vs 13.3 minutes
- Terminal-Bench 4.0: 31% at low vs 60% at high/max — hard coding tasks still benefit from higher effort
- Output tokens per task: 10k at low vs 119k at max
Conclusion: for fast-iterable tasks, multiple low-effort rounds beat a single max-effort run; save max for genuinely hard long-tail problems.
More from coding & agent
- Vercel goes backend: Python hits GA with containers and daemons for async jobs — soleio · 2026-10-16
- WorldBox Boosts Agent Maze Exploration to 230/256 Rooms, 2x the GPT-6 Astra Baseline — ZhitingHu · 2026-10-16
- Isocast: Rust rewrite of OBS boots in 0.7s vs 4s, controllable via CLI and MCP — johnlindquist · 2026-10-16
- Manus partners with OPPO for Find X10 agent mode; Perplexity open-sources SPACE sandbox — heyneighbor · 2026-10-16
- Lindy learns tasks by watching once, adds phone-call features and reshuffles pricing — heyneighbor · 2026-10-16
- Genspark Claw outage blamed on Microsoft, adds iMessage/WhatsApp/Slack/LINE access — heyneighbor · 2026-10-16