💥💥💥 Starting from 2026-July-12, Codex performance dropped 3X ❗️❗️❗️❗️ (Daily data compilation)

What issue are you seeing?

Codex GPT-5.6 Sol daily latency regression

I am seeing a sustained latency regression in Codex when using gpt-5.6-sol.

The charts use daily averages grouped by New York date (EDT). July 12 is normalized to 100%. July 11 and July 19 are partial days because the retained dataset starts at 16:47 EDT on July 11 and July 19 was still in progress when the report was generated.

🔥🔥🔥 performance dropped 3X

<img width="1600" height="700" alt="Image" src="https://github.com/user-attachments/assets/6e282ea8-3262-41a8-87bd-75afc6b4e219" />

🔥🔥🔥 Latency tripled

<img width="1600" height="700" alt="Image" src="https://github.com/user-attachments/assets/b2fea4f6-0984-433d-b1fc-277ce81ae7dc" />

Daily results

| New York date | Successful requests | Mean full-request latency | Median | P90 | Performance index |
|---|---:|---:|---:|---:|---:|
| 2026-07-11 (partial) | 7,849 | 20.62s | 12.40s | 45.21s | 87.1% |
| 2026-07-12 | 37,717 | 17.95s | 11.51s | 32.58s | 100.0% |
| 2026-07-13 | 40,190 | 28.95s | 12.48s | 58.73s | 62.0% |
| 2026-07-14 | 29,577 | 48.84s | 14.50s | 120.05s | 36.8% |
| 2026-07-15 | 26,211 | 56.23s | 13.71s | 149.98s | 31.9% |
| 2026-07-16 | 26,130 | 59.71s | 20.38s | 152.83s | 30.1% |
| 2026-07-17 | 38,157 | 46.40s | 22.13s | 116.51s | 38.7% |
| 2026-07-18 | 30,714 | 46.18s | 17.11s | 104.64s | 38.9% |
| 2026-07-19 (partial) | 22,426 | 59.26s | 26.05s | 154.13s | 30.3% |

Method

Performance index = July 12 daily mean latency / observed daily mean latency x 100

This is a latency-derived throughput proxy, not measured tokens/sec. It includes every successful gpt-5.6-sol text request with a recorded positive full-request duration. Failed requests are excluded. First-response timing is not shown because retained first-chunk logs do not cover the full period.

The live database and migration backups contain no comparable full-request latency records before July 11 at 16:47 EDT. Earlier days therefore cannot be plotted reliably. The partial July 11 sample was 87.1% of the July 12 baseline, so the available evidence does not show that performance was continuously at 100% before July 12.

Summary

• Mean full-request latency increased from 17.95s on July 12 to 59.26s on July 19 so far: approximately 3.30x slower.
• The latency-derived performance index fell from 100% to 30.3%.
• The P90 increased from 32.58s to 154.13s, showing a particularly severe long-tail regression.
• The regression is visible across consecutive daily aggregates, not only isolated samples.

Please investigate whether there is an ongoing Codex / gpt-5.6-sol request and streaming latency regression.

Other

I am not sure if Codex's recent free token reset for so many users has caused server congestion, leading to a massive drop in request performance. I am currently running more than 10 "20x Pro" accounts in parallel to develop different modules for a single project. Even though I pay over $2,000 per month, our project progress has slowed down by a factor of three over the past week.

🤮🤮🤮🤮🤮

I hope you can prioritize guaranteeing that performance does not drop for paying users before giving everyone token reset opportunities. As things stand, even if I wanted to spend more money to buy more tokens, it wouldn't help because the speed ceiling is completely bottlenecked. Because the request speed has dropped threefold, the token consumption rate has become too slow, meaning our current token quota cannot even be used up.

Therefore, I hope you can restore the performance to what it was before July 12th. If some users feel they are consuming tokens too quickly, you could offer them an option to intentionally reduce their response speed by three times (down to the current speed level). This way, everyone can get exactly what they need.

What steps can reproduce the bug?

NA

What is the expected behavior?

_No response_

Additional information

_No response_