news

BudgetPixel's image arena separates its top two models by 6 Elo, and their intervals overlap by 14 points

The October 12 refit puts GPT Image 2.5 Flare at 1613 and GPT Image 2.5 Sunburst at 1607, and ranks Nano Banana 2.1 fourth on 665 votes.

By Daniel Ochoa · 13 October 2026

The image leaderboard at BudgetPixel put GPT Image 2.5 Flare in first place on October 12 with a rating of 1613, six points clear of GPT Image 2.5 Sunburst at 1607. The board also prints 95% intervals for both, 1603 to 1624 for the leader and 1597 to 1617 for the runner-up, which overlap across a 14-point span. By its own arithmetic, the arena has not separated the two models, and both of them come from OpenAI.

40,994 votes, and the top of the board is where they are concentrated

The header states 40,994 votes, 20 models competing, and a rule that a model is ranked after 40 votes in a category. The per-model counts in the top rows sit between 3,816 and 4,049: 3,885 for Flare, 3,910 for Sunburst, 3,816 for GPT Image 2, 3,908 for Nano Banana 2 Lite, 4,000 for Grok Imagine Image 2 and 4,049 for Seedream 5.0 Pro. Since each vote is a pairwise comparison, those per-model counts are appearances rather than votes, and twenty models at roughly 4,000 appearances each is consistent with the 40,994 total only if sampling is close to even. It is not even everywhere.

Fourth place rests on 665 votes

Google's Nano Banana 2.1, which went generally available earlier this month and took its predecessor's slot on the way in, is listed fourth at 1562 on 665 votes, about one sixth of what the entries around it carry. Its interval runs from 1537 to 1585, which contains both GPT Image 2 at 1577 above it and Nano Banana 2 Lite at 1546 below it. Fourth is one of at least three positions the data permits. This is the same failure mode that shows up on video boards, where the published rank range is wider than the gap being claimed, and it is worth reading the interval column before quoting the rank.

Latency, price and failure rate are the columns doing new work

Most public preference boards stop at Elo. This one adds measured speed, a credit price per image and a failure rate, and those columns disagree with the ranking in useful ways. Sunburst sits second on preference while failing 5.4% of the time, roughly one generation in nineteen, against 2.6% for Flare, 0.7% for GPT Image 2 and 0.0% for both Nano Banana 2.1 and Seedream 5.0 Pro. Failures that never reach a voter cannot cost a model preference points, so a rating and a failure rate measure different things and should be read side by side. Generation time spans 10.6 seconds for Nano Banana 2.1 to 59.3 seconds for GPT Image 2, which the board also lists as the most expensive of the group at $0.060 an image while rating third.

For comparison, our own benchmark scores models on named capabilities rather than on a single overall preference number.

The preference leader is fifth by request volume

OpenRouter's image model rankings, covering usage through the same day, put Google's Nano Banana 2 first at 1.37 million requests and up 10%, Sunburst second at 1.09 million and down 6%, Nano Banana 2 Lite third at 716,000 and up 15%, Meta's Muse Image fourth at 635,000 and up 33%, and Flare fifth at 620,000 and down 16%. Request share through one router answers a different question than blind preference does, but the model rated highest on October 12 is not the model being called most, and its volume is falling while the Google entries climb.

Two boards refreshed on the same day now describe the image field in incompatible terms. A separate arena running on 11,311 blind votes across 10 models puts GPT Image 2 on top; this one puts GPT Image 2 third, behind two variants that did not exist in that sample. Neither is wrong on its own data. The useful habit is to check the vote count, the interval and the failure rate before treating any of these orderings as a result.

Sources: BudgetPixel AI Image Arena, OpenRouter Image Model Rankings.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.