OpenAI dropped its latest image generation model, ChatGPT Images 2.5, on September 8, and it didn’t take long for the results to speak for themselves. The Sunburst variant, built for high-precision output, climbed to the number one ranking across Arena’s text-to-image, single-image edit, and multi-image edit leaderboards within hours of going live.
To celebrate (or stress-test, depending on your cynicism level), Arena is giving everyone free access to Sunburst for 72 hours. No credit card, no waitlist, just open the platform and start generating.
What Sunburst actually does differently
ChatGPT Images 2.5 ships with two API variants. Sunburst is the precision-focused model, optimized for lighting fidelity, prompt adherence, and subject preservation. Flare is the speed demon, delivering up to 50% lower latency compared to previous models.
As of September 9, Sunburst held an estimated ELO score of approximately 1421 on Arena’s text-to-image leaderboard. For context, Arena uses a competitive ranking system similar to chess ratings, where models are pitted against each other in blind matchups and users vote on which output looks better. An ELO of 1421 puts Sunburst meaningfully ahead of its predecessor GPT-Image-2 and every other model currently on the platform.
Flare positioned itself closely behind Sunburst in those same rankings, which means OpenAI effectively holds the top two spots on the leaderboard right now.
How Arena’s ranking system works
Arena.ai has become the de facto proving ground for AI image generation models. The platform works by showing users two outputs side by side, generated by different models from the same prompt, without revealing which model made which image. Users then vote on which they prefer.
Pricing and the competitive landscape
Once the free trial window closes, Sunburst’s API pricing lands at $5 per million input tokens and $30 per million output image tokens.
The improvements OpenAI highlights in this release, specifically lighting fidelity, prompt adherence, and subject preservation, address three of the most persistent complaints about AI image generation.
Multi-turn editing fidelity is another notable upgrade. This refers to the model’s ability to maintain consistency across multiple rounds of edits to the same image.