Luna: radical performance regression at high effort in a newer Codex CLI

Summary

I run a coding benchmark (105 planted bugs across two real production repos, one agentic run per repo) through codex exec, using GPT-5.6 Luna on a ChatGPT-account auth, across effort tiers from low to max. Each run is a long multi-turn agent session.

Between two Codex CLI builds, Luna's high-effort runs regressed radically in time and billed cost, and in the worst case stopped converging at all. The correctness of completed runs was unaffected. I did not see the same behavior on GPT-5.6 Sol or Terra in my runs.

Environment

• Codex CLI (old, clean): OpenAI ChatGPT extension 26.727 (late July 2026)
• Codex CLI (new, regressed): extension 26.820.71523, codex-cli 0.150.0-alpha.8 (August 2026)
• Auth: ChatGPT account
• Model: gpt-5.6-luna
• OS: Windows

What I measured

On the old CLI (26.727), a full Luna max-effort run completed cleanly:

• about 85 minutes wall clock
• about $1.80 (list-rate estimate)
• about 64M billed input tokens (dominated by cached context re-reads across turns)

On the new CLI (26.820.71523):

• Luna xhigh (one tier below max): about 135 minutes, about $2.50, about 83M input tokens. A lower effort tier now costs more time and tokens than the old CLI's max.
• Luna max (a re-run of the same benchmark): it failed to converge. It ran 3 hours 23 minutes on a single repo leg out of two, produced a 54 MB session log and hundreds of millions of re-processed context tokens, and never produced a result, so I aborted it.

Likely mechanism

The new build appears to auto-compact the context around 200K tokens. Luna's usable context runs past 272K (Luna bills a 2x input and 1.5x output premium above 272K per the published rates). On long agent sessions the CLI seems to compact before the model's real limit and then re-process context in a loop, which inflates billed input tokens and wall clock.

Impact

• Time and billed cost. Roughly doubled at xhigh, and effectively unbounded at max where the run did not finish.
• Run non-convergence in the worst case.
• Not answer quality. Completed runs scored normally. Luna's fix-rate stayed monotonic across effort: 4, 9, 13, 23, 33 out of 105 for low, medium, high, xhigh, max.
• Only the high-effort tiers (xhigh, max) are affected. Low and medium runs are short enough that they do not reach the compaction threshold.

Expected

High-effort Luna runs on the new CLI should complete in comparable time and cost to the old CLI, not regress by 2x or more, and not fail to converge.

I can share the more on request.

Pawel Huryn

Reproduce the regression through codex exec using the reported old and new CLI builds, gpt-5.6-luna, high-effort tiers, and the same benchmark on Windows. Compare session logs, context compaction near 200K tokens, wall time, and billed input tokens; done means high-effort runs complete without runaway reprocessing and return to comparable time and cost.