Around June 16, 2026, ChatGPT Plus users started reporting a severe regression in Codex rate-limit consumption. The same model, same configuration, same account — yet the daily prompt count dropped from 20+ to just 2–3 before hitting the 5-hour cap.
The most detailed report is Issue #28879 on the openai/codex repository.
On GPT-5.5 (Plus plan), the per-token rate-limit cost appears to have increased roughly 10–20x. A 57K-token prompt that consumed ~1% of the 5-hour budget on June 12 now translates to 10–27% for a 20K-token prompt. The cause is unconfirmed, but the timing matches an active incident on the OpenAI status page.
What Happened
Reporter @mihneaptu has been using the Codex desktop app on Windows 11 with gpt-5.5. Configuration remained unchanged across the regression window.
June 12 (normal — 20+ prompts/day)
A 57K-token prompt with 3.6K reasoning consumed only ~1% of the 5-hour budget.
June 18 (broken — 2–3 prompts/day)
A 20K-token prompt with near-zero reasoning now consumes 10–27% of the 5-hour budget. That's roughly a 10–20x increase in per-token rate-limit cost.
What Was Ruled Out
The reporter verified from session logs:
- Not reasoning effort: Reasoning tokens dropped from 3,645–5,125 to 0–152. The issue is not increased thinking
- Not prompt bloat: Input tokens decreased from 57K–80K to 18K–23K. Prompts are actually smaller
- Not config change: model_reasoning_effort was high/xhigh on both good and bad days
- Not the weekly cap: Secondary (7-day) window was at only 12%. The problem is solely in the 5-hour primary window
Timing Correlation
The OpenAI status page shows an active incident starting around June 16 — matching exactly when the usable prompt count dropped from ~9 sessions/day to ~4.
Issue #28879 suggests either the Plus-tier GPT-5.5 rate-limit budget was reduced or the per-token weighting was changed. As of now, there's no official response from OpenAI.
Impact Assessment
This is an unresolved incident with no official statement from OpenAI yet. If you depend on Codex rate limits for critical workflows, consider diversifying across multiple models and providers.
References
- Issue #28879: https://github.com/openai/codex/issues/28879
- codex-reset (CLI tool for banked resets): https://github.com/aaamosh/codex-reset
- OpenAI Status: https://status.openai.com
Summary
Codex Issue #28879 strikes at the core of the Plus plan's value proposition. A 10–20x increase in per-token rate-limit cost — whether a bug or an intentional change — makes GPT-5.5 nearly unusable for anything beyond a handful of interactions per day.
Affected users should manage their banked resets with codex-reset and consider spreading their workload across alternative models like Kimi K2.7 Code or Claude Opus while OpenAI investigates.
