OpenAI Codex changelog

MarginLab Codex tracker shows GPT-6 Sol pass rates falling while its baseline is still being collected

MarginLab's Codex tracker (GPT-6 Sol at high effort, from September 24, 2026) showed 54% on the latest day (50 tasks), 59% over 7 days (350 tasks) and 79% over 30 days (1,450 tasks). The page marks 1-day (-29.4%) and 7-day (-24.9%) changes as statistically significant but says degradation detection is paused while the new baseline is built.

Who is affected:

Codex CLI users on GPT-6 Sol at high effort.

What to do:

Treat the drop as provisional and compare GPT-6.1 Sol on your own tasks.

OpenAI fixes an image encoding bug that degraded GPT-6 Sol and Luna image understanding

OpenAI's API changelog says a bug in image encoding degraded image understanding in GPT-6 Sol and GPT-6 Luna, and the fix improves visual tasks in the API and Codex, including computer use.

Who is affected:

Apps and Codex sessions that sent image inputs or used computer use with GPT-6 Sol or GPT-6 Luna after their September 22 release.

What to do:

Rerun evaluations and retry workflows that used image inputs with GPT-6 Sol or GPT-6 Luna.

Codex users report GPT-5.5 reasoning tokens clustering at 516, linked to weaker answers

A GitHub issue analyzing Codex token metadata from February 1 to June 27, 2026 found gpt-5.5 responses landing disproportionately at exactly 516 reasoning tokens, consistent with a reasoning-budget threshold. The author says it does not prove truncation. The thread drew 372 points and 152 comments.

Who is affected:

Codex users running gpt-5.5 on complex tasks.

What to do:

No action needed. GPT-5.5 retired from Codex with ChatGPT sign-in on October 14, 2026.

OpenAI fixes Codex bug that burned usage limits too fast

OpenAI's Codex lead said auto-review and subagents sometimes ran more often than intended or retried too aggressively, draining usage limits. Fixes were deployed and limits were reset for all users.

Who is affected:

Codex users who hit usage limits early in late June 2026.

What to do:

No action needed. Limits were reset.

OpenAI publishes its Codex degradation investigation and finds smaller issues, not one large cause

After public reports that Codex performed worse around the gpt-5-codex launch on September 15, 2025, the Codex team investigated full time, removed over 60 feature flags, and reported a combination of behavior shifts (such as compaction) and smaller concrete issues rather than one large regression.

Who is affected:

Codex CLI and cloud users who saw weaker results in September and October 2025.

What to do:

Update the Codex CLI and start new conversations for targeted tasks instead of relying on repeated compaction.