Three weeks after its launch, ChatGPT Images 2.5 has taken the lead in both image generation and editing rankings on Arena, a platform that ranks text-to-image models based on public votes. The new model, introduced on September 8 through ChatGPT and its API, has two variants—Flare and Sunburst—that outperformed its predecessor, GPT Image 2, in the September edition of the rankings. This makes OpenAI, the company behind ChatGPT, the current leader in the field. However, not all models saw improvement: some previously strong models lost ground, while others remained stable. On the API side, OpenAI offers two versions of ChatGPT Images 2.5. Flare, the default option, reduces the time it takes to generate images by half compared to GPT Image 2. Sunburst, which is slower, is designed for more precise creative tasks, such as detailed image editing. Sunburst took the top spot, while Flare came in second, pushing GPT Image 2 to third. In August, other models like MAI-Image-2.6 and Grok Imagine 2.0 were in the top 10, but they had fewer votes. Microsoft's MAI-Image-2.6 stayed mostly the same, only dropping two places due to new competitors. SpaceXAI's Grok Imagine 2.0, however, lost 15 points as more votes accumulated, falling from third to sixth. Meta’s Muse Image dropped two spots to seventh, and Reve kept both of its models in the top 10. Google only had Nano Banana 2 in the rankings, while Alibaba and MAI-Image-2.5 fell out of the top 10. In the lower ranks, Qwen Image 3.0 Pro and Seedream 5.0 Pro tied for 11th, and Microsoft's model was 12th. The top 10 image generation models in September 2026, based on Elo scores, are: GPT Image 2.5 Sunburst (OpenAI) with 1,424 points, GPT Image 2.5 Flare (OpenAI) with 1,401, GPT Image 2 medium (OpenAI) with 1,383, MAI-Image-2.6 (Microsoft) with 1,335, Reve 2.1 (Reve) with 1,302, Grok Imagine 2.0 low (SpaceXAI) with 1,301, Muse Image (Meta) with 1,276, Reve 2.0 (Reve) with 1,269, Nano Banana 2 web search (Google) with 1,261, and Seedream 5.0 Pro (ByteDance) with 1,256. In the image editing category, Sunburst widened its lead. This ranking evaluates how well models can modify an existing image based on a prompt. Sunburst scored 1,526 points, outperforming Flare by 44 points and GPT Image 2 by 65. This aligns with OpenAI's design goal for Sunburst, which is intended for tasks requiring fine control over image details. Grok Imagine 2.0 and MAI-Image-2.6 took fourth and fifth places, tied closely. Muse Image followed, while MAI-Image-2.5 dropped behind. Google placed Nano Banana Pro ninth, and Reve’s models fell to fourteenth place. The September rankings included 56 models and over 29.9 million votes, compared to about 6.5 million for text-to-image generation. The top five image editing models were: GPT Image 2.5 Sunburst (OpenAI) with 1,526 points, GPT Image 2.5 Flare (OpenAI) with 1,482, GPT Image 2 medium (OpenAI) with 1,461, Grok Imagine 2.0 low (SpaceXAI) with 1,430, and MAI-Image-2.6 (Microsoft) with 1,429. The Arena rankings are determined by anonymous duels where users submit prompts and then choose between two images generated by different models, without knowing which model created each. Each vote updates an Elo-type score, where beating a higher-ranked model earns more points, and losing to a lower-ranked one costs more. Confidence intervals and possible rank ranges are provided to show the uncertainty due to the number of votes. Newer models with unstable scores are marked as "preliminary."