Meta-Reasoning improves tested long-horizon agent results
GPT-5.5 and Opus 4.8 results are separate model comparisons; they are not endpoints of one budget curve.
The Meta-Reasoning paper reports 71.5% with GPT-5.5 versus its Codex baseline at 58.0%, and 67.2% with Opus 4.8 versus its Claude Code baseline at 65.5%. Each comparison has its own model and baseline. Comparing 71.5% with 67.2% does not show diminishing returns as tokens increase.
The paper reports continued improvement over its tested budget range where baseline agents plateau. This is an empirical result within those experiments, not a promise of indefinite scaling. The original article reversed that trend and included unverified claims about a ten-kilobyte run account and an unrelated AsyncLLM benchmark; those claims have been removed.
For a deployment decision, reproduce the model-specific accuracy/cost curve, including controller calls and the stopping policy. Keep the tested horizon and budget attached to the results rather than claiming a controller is always cheap or always expensive.
Editorial correction (2026-10-13 UTC): Corrected cross-model comparisons and the tested budget trend.