Same here, it matches what we’ve been measuring almost exactly
We hit the same thing on POST /v1/images/edits with gpt-image-2. It started
for us on 2026-08-09 as well, with no change on our end.
A controlled experiment that isolates where it breaks
Four series inside the same 15-minute window (2026-08-10, 09:27–09:31 UTC), same
egress, same Cloudflare PoP (BRU), byte-identical 137 KB multipart bodies (1024 px JPEG + PNG mask, synthetic test images):
| # | Setup | Result |
|---|---|---|
| A | postman-echo.com/post (also Cloudflare-fronted) |
15/15 → 200 |
| B | /v1/images/edits, invalid key, gpt-image-2 |
15/15 → 401 + x-request-id |
| C | /v1/images/edits, real key, nonexistent model |
10/10 → 400 + x-request-id |
| D | /v1/images/edits, real key, gpt-image-2 |
10/10 → 520, no x-request-id |
B and C are the interesting ones: the exact same 137 KB body reaches OpenAI’s application layer intact and comes back with their own error codes and their own
request IDs in 0.2–0.6 s. The only variable that changes in D is whether image
generation is actually attempted.
We also ruled out an inline TLS-inspecting proxy — by measurement
Unlike you, our egress does pass through a corporate TLS-inspecting forward proxy, so support quite reasonably asked whether that could be disrupting long-lived connections. It isn’t. Five non-streaming POST /v1/chat/completions
calls (gpt-5.5, 4000 max completion tokens), where the connection sits idle for
over a minute waiting, over the same proxy and the same PoP:
09:42:01 200 79.1 s openai-processing-ms 78530
09:43:19 200 77.8 s openai-processing-ms 77297
09:44:38 200 78.6 s openai-processing-ms 78316
09:45:54 200 75.8 s openai-processing-ms 75500
09:47:14 200 79.5 s openai-processing-ms 79205
5/5 succeeded at ~78 s each. Every failing image edit died between 0.33 s and 25.9 s — all shorter. So connection longevity isn’t the variable. And your report closes the loop from the other direction: you have no proxy at all and see the identical failure. Between the two of us, the client-side network is out.
CF-Ray IDs for series D (all 520, 2026-08-10, times are response completion, UTC)
09:28:36 a28df87019fd985a-BRU 7.7 s
09:28:43 a28df8a1dcff1dfe-BRU 7.3 s
09:28:44 a28df8d0da651037-BRU 0.33 s
09:28:51 a28df8d43e51b9ba-BRU 7.1 s
09:29:04 a28df9023edd1327-BRU 12.7 s
09:29:10 a28df95308421327-BRU 5.6 s
09:29:28 a28df9776e0e13d7-BRU 18.1 s
09:29:29 a28df9ea2e3daeaf-BRU 0.53 s
09:29:55 a28df9eedc18df65-BRU 25.9 s
09:30:21 a28dfa92beceb9b9-BRU 25.9 s
Note the 0.33 s and 0.53 s ones — same latency as our successful 400 responses.