Codex
gstack now runs its eval cases on Codex every Monday, beside Claude
claude-opuscodex-solgstack's paid eval cases can now run on Codex, each one through a full Codex runtime. Every Monday the census runs two arms side by side: Claude Opus 5.5, and Codex on gpt-6.1-sol.
Pilot
On its first run, Codex passed 48 of 52 gate cases at a sixth of the cost
On its first pilot, Codex cleared 48 of the 52 gate cases with no retries. The whole run cost about a sixth of the same cases on Claude Fable.
Claude-only
Claude-only cases skip on Codex and never count as failed
Cases that need Claude are skipped on Codex with a recorded reason, never marked failed. The 26 rework cases stay Claude-only, and no threshold or Claude verdict changed.
Monthly
Once a month the census adds Claude Fable and gpt-6-astra
Once a month, the census adds Claude Fable and Codex on gpt-6-astra. Codex runs now report a real dollar cost, and reports name the host whenever it is not Claude.
Try it
Running one case on Codex yourself: the release notes, “If you…”
CHANGELOG.mdgstack 1.91.149.0 release notesTo run a single case on Codex yourself, follow the If you section of gstack's release notes. Which agent do you run gstack in?
















