Summary
Starting ~June 16, my ChatGPT Plus Codex budget drains in 2–3 prompts on gpt-5.5, where the same model/plan/app previously gave 20+ prompts. My Codex session logs (token_count / rate_limits events) show the limit-% consumed per token has increased roughly 10–20×, with no change on my side — prompts are actually smaller now and reasoning tokens are near zero.
Environment
- Codex desktop app, Windows 11
- Model:
gpt-5.5,service_tier = "default" - Plan:
plus(confirmed inrate_limits.plan_type) - Config unchanged across the regression window
Evidence — same plan, same model, same app
June 12 (working — 20+ prompts/day):
| Prompt | Input tokens | Reasoning tokens | Primary limit % |
|---|---|---|---|
| 06:08 | 22,334 | 488 | 2% |
| 06:09 | 52,161 | 1,130 | 3% |
| 06:51 | 57,112 | 3,645 | 9% → 11% |
| 06:54 | 79,961 | 5,125 | 15% → 16% |
→ ~57K-token prompt with 3.6K reasoning ≈ 1% of the 5h budget
June 18 (broken — 2–3 prompts):
| Prompt | Input tokens | Reasoning tokens | Primary limit % |
|---|---|---|---|
| 07:22:27 | 18,910 | 152 | 18% |
| 07:22:30 | 20,480 | 0 | 45% |
| 07:22:34 | 23,469 | 48 | 57% |
| 07:22:43 | 23,631 | 0 | 68% |
| 07:22:50 | 23,839 | 0 | 77% |
→ ~20K-token prompt with zero reasoning ≈ 10–27% of the 5h budget
Per-token limit cost changed ~10–20×. Smaller prompts, near-zero reasoning, yet each consumes 10–20× more of the 5-hour budget than a 57K-token xhigh prompt did on June 12.
What I ruled out (verified from logs)
- Not reasoning effort: reasoning tokens are 0–152 today vs 3,645–5,125 on the good days.
- Not prompt/context bloat: input tokens dropped from 57K–80K to 18K–23K.
- Not my config:
model_reasoning_effortwas alreadyhigh/xhighon the working days (June 12–15). xhigh gave 30–40+ prompts on June 3, 12, 13, 15. - Not the weekly cap: secondary window is only at 12%. The 5-hour window is the one draining.
Corroborating timing
OpenAI status page shows an active incident ongoing ~2 days (since ~June 16), which matches exactly when my usable prompt count dropped from ~9 sessions/day to ~4.
Ask
Was the Plus-tier gpt-5.5 rate-limit budget reduced or the per-token weighting changed around June 16? If intentional, what's the new budget? If not, can it be investigated?
Session log references
- Good day:
rollout-2026-06-12T09-07-36-019eba71-5c57-70d0-a8fa-aa999731e9ff.jsonl - Bad day:
rollout-2026-06-18T10-22-14-019ed99b-ea9d-7fa1-a86b-dca4e55a8e3e.jsonl - Rate-limit data is in
token_countevents withrate_limitspayload (fields:primary.used_percent,secondary.used_percent,plan_type,resets_at).