The State of AI Image Models, September 2026

OpenAI just closed the speed gap it had been losing on

GPT Image 2.5 Flare produces an image in 12.4s median, against 35.9s for GPT Image 2 — a 2.9× speed-up in a single version. That puts it level with Google's Nano Banana 2 at 12.3s, a difference well inside the run-to-run noise on both sides.

When we measured in June, OpenAI's image model was several times slower than Google's, and that gap was the most striking number in the report. It is gone. After a year of Google owning the speed argument, the two are now indistinguishable on it.

Time-to-image, by model

Four fixed text-to-image scenes, three runs per model per scene, one generation per request.

Method: wall-clock time from request to returned image, identical scenes, one shot per run, no retries counted. GPT Image 2.5 was measured on 9 September 2026; GPT Image 2 and Nano Banana 2 on the same scenes through the same harness on 10 September. A fifth image-editing scene was run but is excluded here, because Nano Banana 2 is not served on the image-edit endpoint and it would not have been like-for-like. Every row has at least 12 runs behind it.

What that looks like for a real user

Benchmarks run one clean request at a time. Real users hit queues, retries and busy hours. So we also measured every image generation on PromptFrenzy over the last 60 days — all prompts, real traffic:

The medians track the benchmark closely, which is a good sign the lab numbers aren't flattering anyone. But the means sit far above the medians on both models: a minority of slow generations drags the average up, and on GPT Image 2 the slowest tenth take 80 seconds or more. The typical wait is better than the average number suggests.

GPT Image 2.5 is absent here because it is not yet serving production traffic on our platform, so we have no production sample for it. Internal, admin and automated-test accounts are excluded throughout.

The field is moving faster than the models render

Speed matters more this year because the roster keeps growing. Of the AI image and video models we track with a known launch date, a third shipped in the first half of 2026 alone — Google, OpenAI, xAI, Black Forest Labs and ByteDance all pushed new models in the last six months.

Same prompt, every model

The most useful thing we publish isn't a number — it's the side-by-side. We take one brief and run it through every model, one shot each, no cherry-picking. The four scenes behind the table above are deliberately awkward: a refraction test, a reflection test, a compound-instruction test, and an open-ended four-panel comic. They are built to separate models rather than flatter them, and the differences in interpretation, text rendering and physical plausibility rarely match each model's reputation.

What it means

• →The speed argument is over, for now. OpenAI and Google are level on time-to-image. Anyone still choosing between them on speed is working from last quarter’s data.
• →Version jumps matter more than vendor reputation. GPT Image 2 and GPT Image 2.5 are further apart on speed than GPT Image 2.5 and its closest competitor.
• →Averages hide the wait. Median and 90th-percentile latency tell a user far more than a mean, and almost nobody publishes them.
• →There is still no single “best.” The right model depends on the task — which is the whole reason to compare on your own prompt rather than trust a leaderboard.