OpenAI API still burns reasoning tokens during context compaction even with effort set to none
mhmazur · x · 2026-09-04
While testing the new OpenAI Provider Adapter harness on ARC-AGI-3, the author found that setting reasoning effort to "none" in the API request doesn't stop reasoning tokens from being consumed during context-window compaction. These tokens reflect the model's internal work of deciding what context, state, and prior reasoning to retain — not reasoning about which game action to take.
Another gotcha: with a non-none reasoning level, automatic compaction triggers don't split reasoning token usage between compaction and action decision in the usage data. Mostly an edge-case issue, but real visibility loss for anyone doing deep token-level analysis of agent runs.
More from Models
- LMArena launches Image Edit leaderboard to rank models — arena · 2026-09-05
- Knowledge worker seeks cheaper Claude alternatives with strong instruction-following and long context — SemiMagnum · 2026-09-05
- Every to host live camp comparing Anthropic Fable 5.1 and OpenAI GPT-6 Astra — danshipper · 2026-09-05
- Real-world testing suggests Artificial Analysis Index is gamed and unrepresentative — PerformanceRound7913 · 2026-09-05
- Microsoft's MAI-Image-2.6-Flash hits #3 in image editing, jumping 34 Elo over last Flash — ArtificialAnlys · 2026-09-05
- User claims Artificial Analysis Index is easy to game, doesn't match real-world performance — PerformanceRound7913 · 2026-09-05