Across 6,198 image generations run through AIGE's web studio between March and August 2026, 92.6% delivered an image. Leave out the jobs that failed because of an outage on our side and the figure is 93.9%. The rest failed for four reasons: the model returned an error (45% of failures), a content-safety filter blocked the request (19%), our platform refused the job (19%), or the request itself was invalid (18%).
The models differ more than the average suggests. Recraft V4.1 delivered 99.1% of the time once our own outages are excluded, with a median of 15 seconds. Nano Banana Lite was the fastest at 8 to 9 seconds. GPT Image 2 edits were blocked by content policy more often than any other model: 67 of 1,040 runs, or 6.4%.
Success rate and speed by model
Every row is a model family as AIGE's web studio called it. Versions changed during the period (for example Nano Banana 2 to 2.1), so treat each row as the family, not one version. "Delivered" counts every run; "delivered when we were up" leaves out runs our platform refused. Time is from submitting the job to the finished image.
Model | Runs | Delivered | Delivered when we were up | Median time | 9 in 10 within | Most common failure |
|---|---|---|---|---|---|---|
Nano Banana (edit) | 2,518 | 94.2% | 94.4% | 63 s | 94 s | Model error |
GPT Image 2 (edit) | 1,040 | 90.2% | 90.9% | 93 s | 182 s | Content policy |
Recraft V4.1 | 727 | 95.5% | 99.1% | 15 s | 17 s | Our outage |
Nano Banana (new image) | 599 | 91.2% | 91.5% | 32 s | 63 s | Model error |
GPT Image 2 (new image) | 399 | 90.7% | 94.5% | 62 s | 121 s | Our outage |
Nano Banana Lite (edit) | 307 | 87.6% | 92.4% | 9 s | 19 s | Invalid request |
Nano Banana Pro (edit) | 249 | 92.8% | 93.5% | 35 s | 95 s | Model error |
Nano Banana Lite (new image) | 188 | 85.6% | 89.9% | 8 s | 9 s | Invalid request |
Recraft V4 Pro | 96 | 96.9% | 96.9% | 62 s | 63 s | Model error |
Nano Banana Pro (new image) | 75 | 98.7% | 100% | 31 s | 39 s | Our outage |
Why AI image generations fail
We sorted all 459 failed image runs by the error message the provider or our server recorded.
The model returned an error (205 runs, 45%). The provider accepted the job and then returned no image or an "unprocessable" response. Nano Banana edits account for most of these. Running the same request again usually works.
Content-safety blocks (86 runs, 19%). The model's own filter refused the prompt or the input picture. GPT Image 2 edits carry 67 of the 86. Rewording the prompt, or switching to another model for that shot, is the fix.
Our platform refused the job (87 runs, 19%). These came from short outages on our side, when our account with the model provider was paused. They are not the model's fault, which is why the table shows a second column without them.
Invalid requests (81 runs, 18%). An unsupported aspect ratio, too many reference images, or a missing field. Nano Banana Lite was the most sensitive to these.
What this means when you pick a model
For speed: Nano Banana Lite and Recraft V4.1 returned most images in under 20 seconds.
For edits of an existing picture: Nano Banana edits delivered 94% of the time; GPT Image 2 edits were slower and more often blocked by its safety filter.
For a reliable first draft: Recraft V4.1 failed least once our own outages are set aside.
For cost: prices differ by model and size; see the AI image generator page and the Nano Banana guide for credits per image.
What happens to your credits when a generation fails is covered in why AI generations fail and what happens to your credits.
How we counted
Source: AIGE's own task log for the web studio, read on 26 October 2026: 6,451 image and video jobs from 2 March to 26 October 2026, each with its final status, timestamps and, for failures, the recorded error message. No prompts or personal data were read.
Who ran them: AIGE's customers and AIGE's own team, including testing. We cannot separate the two in this log.
Period: almost all runs are from March to August 2026. Since September the web studio records generations in a different log, so recent months are not in these numbers.
Video: the log holds too few video runs per model (under 60 each) to publish rates for Kling, Veo or Seedance.
Models with fewer than 20 runs are left out.
Quick answers
Question | Answer |
|---|---|
How often does AI image generation fail? | In AIGE's log, 7.4% of image runs failed (6.1% leaving out our own outages). |
Which AI image model is most reliable? | Recraft V4.1: 99.1% delivered when our platform was up, across 727 runs. |
Which AI image model is fastest? | Nano Banana Lite: a median of 8 to 9 seconds. |
Why was my AI image blocked? | The model's content-safety filter refused the prompt or input picture. GPT Image 2 edits did this most often (6.4% of runs). |
Comments
Be the first to comment.