Skip to main content
  • Published: November 20, 2026
  • Last checked: 11:35 on November 20, 2026 (UTC+8)
  • Status: Monitoring. Every -vip model fails more often than usual; the web route gpt-image-2.6-web and GPT Image models on official-forward groups are working normally
This morning’s advice to switch to a Sunburst VIP model no longer avoids the errors. Retry a failed request once in your code; if a 1K or 2K image fails again, resend it to the web route with the same key.

Update at 11:35 on November 20: every VIP model is affected

From 10:00 to 11:35, about 48% of requests across all VIP models returned an image, dropping to about 34% around 10:30. Most failures are a 504 after waiting about 5 minutes or an immediate 503, with some 502s. Failed requests are not charged. Over the same period gpt-image-2.6-web succeeded on about 99.7% of requests, and GPT Image models on official-forward groups kept returning images. Handle errors like this: For complete code that falls back to the web route automatically, see Fall back to the web route when VIP fails. Keys in the GPT Image VIP group can only call VIP models; use a default group key for the web fallback.

Success rate by VIP model earlier on November 20

Window: 00:00 to 10:00 on November 20, 2026 (UTC+8), when the Sunburst VIP models were still largely fine; for 10:00 onward, follow the update above. Success rate is requests that returned an image divided by those plus requests that failed with a server error. Invalid parameters and moderation rejections are not counted.

Move from gpt-image-2-vip to a Sunburst VIP model

Since 10:00 the Sunburst VIP models are affected too, so changing models does not replace the retry and web fallback above. Once VIP recovers, use these steps if you want to move off gpt-image-2-vip:
  1. Change model to one of the Sunburst VIP models below. Your key, the endpoints (/v1/images/generations, /v1/images/edits), and the other parameters stay the same.
  2. Send a few requests first and open the saved images to check size and style. 2.5 and 2.6 are newer model versions, so the same prompt will look somewhat different from 2.0.
  3. Confirm the model name and charge in the usage log, then move the rest of your traffic.
Minimal change for a generation request:
For edits, change only the form field: -F "model=gpt-image-2.5-sunburst-vip". To save the returned images, follow the GPT Image 2.5 VIP examples.

Who is affected

  • Affected: generation and edit requests to every -vip model: gpt-image-2-vip since November 19, and the 2.5 and 2.6 Sunburst and Flare VIP models since 10:00 on November 20. Failures are mostly 5xx errors (usually 502, 503, or 504), occasionally 429, and are not charged.
  • Not affected: the web route gpt-image-2.6-web and GPT Image models on official-forward groups.

FAQ

Will the VIP models be removed?

No. gpt-image-2-vip and the other VIP models stay available at the same price, and this page will be updated when they are stable again. For batch jobs that need a high success rate, add the retry and web fallback to your code now.

How does the web route differ in price and output?

The price is the same: gpt-image-2-vip, gpt-image-2.5-sunburst-vip, gpt-image-2.6-sunburst-vip, and gpt-image-2.6-web all cost $0.03 per image, for generation and edits alike. See Models and pricing for current prices. The output differs: the web route renders natively up to about 2K, cannot select Flare or Sunburst, and 4K sizes carry about 2K worth of detail. Parameters are in GPT Image 2.6 web route.

After VIP recovers, which Sunburst VIP model should I pick?

Pick gpt-image-2.5-sunburst-vip if you need xhigh or max quality. Pick gpt-image-2.6-sunburst-vip for animated GIFs or the newest version. Both render natively from 1K to 4K; parameters are in GPT Image 2.6 VIP per image.

How do I cut failures in my code?

On a 429 or 5xx error, wait 30–60 seconds and retry once. If it fails again and size is 2K or smaller, resend the same request to gpt-image-2.6-web; for 4K, return the error or use official forwarding. Lower your concurrency as well. The gateway does not switch models for you, so the fallback has to live in your code; see Fall back when VIP fails for a pattern.

What if I get an image full of noise or corrupted blocks?

Find the request ID in the usage log and open a ticket with it. We check each one.