GPT Image 2.5 登顶图像竞技场双榜

OpenAI’s GPT Images 2.5 takes the top two spots on the Artificial Analysis Image Arena, generating higher quality images in less than half the time, while keeping the same price as GPT Image 2 Nearly five months after GPT Image 2’s release, OpenAI has released GPT Image 2.5 Flare and Sunburst. OpenAI positions Flare as the model choice for most applications, designed for everyday image generation at lower latency than GPT Image 2. Sunburst is positioned by OpenAI as offering higher quality for detailed creative work and tighter control across edits, but taking longer to generate images. Both models handle Text to Image and Image Editing. Flare and Sunburst take the top two spots on both Artificial Analysis image leaderboards, with Flare #1 in Text to Image and Sunburst #1 in Image Editing. Our evaluation of both models was performed at “max” quality. The largest gains were in Image Editing, with the biggest improvements over GPT Image 2 in: ➤ Composition & Framing: camera changes, zoom-outs and layout changes that must keep the scene intact, where Sunburst now holds the widest lead of any editing action. ➤ Text or Symbol Edits: changes to words, numbers, logos and symbols within an existing image, where Sunburst's gain over GPT Image 2 is second only to Composition & Framing. These gains come at the same estimated output price as GPT Image 2 (high) for 1024 × 1024 images at the evaluated quality settings. Both models are available through the OpenAI API. OpenAI has also released ChatGPT Images 2.5 in ChatGPT, ChatGPT Work, and Codex. Congratulations to @OpenAI on the release! See below for our analysis and example outputs of GPT Image 2.5 Flare and Sunburst in the Artificial Analysis Image Arena 🧵

GPT Images 2.5 are a material improvement in quality at the same price as GPT Image 2 OpenAI prices both new models at the same token rates as GPT Image 2: $30 per million image output tokens. This makes it the clear choice for the highest quality images, though the next model on the frontier, Microsoft's MAI-Image-2.6, costs a fifth as much at $38.90 per 1k images. Meta's Muse Image costs $10 per 1k, a twentieth of the price, and holds a top-seven spot on both boards.

Flare pairs its quality gains with a much shorter wait for images and edits, with GPT Image 2.5 Flare (max) having a ~63% faster latency than GPT Image 2 (high) In our independent inference speed benchmarking, our preliminary results show similar improvements to OpenAI's claim of 50% lower latency for Flare versus GPT Image 2: ➤ Flare (max) reduced median generation time by 63%, from 142.9 to 52.7 seconds ➤ Sunburst (max) reduced median generation time by 24% to 108.9s Our testing used 1024 × 1024 PNG outputs and represents a median across several testing runs including inference and image transfer.

GPT Image 2.5 Sunburst is the leading model in 8 of our 10 Image Editing use cases, with the largest leads in Frontier, Productivity & Knowledge Work, and Social Media & Creator Content. ➤ Frontier covers stress-test prompts that today's best models still get wrong: extreme compositional complexity, many precise simultaneous constraints, and hard physics or reasoning. ➤ Productivity & Knowledge Work covers diagrams, infographics, charts, slides, educational explainers, and other document and workplace visuals. ➤ Social Media & Creator Content covers content made to post: video thumbnails, promo cards, and channel and profile art.

In Image Editing, GPT Image 2.5 Sunburst sets the frontier in 6 of 7 editing actions and 8 of 10 use cases Our Image Editing arena scores models on two axes: seven editing actions which represent what an edit asks the model to do, and the same ten use cases as Text to Image. Before this release, the seven Image Editing actions were split between five different leaders. GPT Image 2.5 Sunburst now leads six of them. Sunburst's largest leads over the next-best model are in Composition & Framing and Object-Level Edit, both over GPT Image 2 (high), and Text or Symbol Edits, over Microsoft's MAI-Image-2.6. Sunburst closely trails Microsoft's MAI-Image-2.5-Pro in Enhancement & Restoration, the one editing action it does not lead. ➤ Composition & Framing covers camera changes, zoom-outs, and layout changes that must preserve the scene or design. ➤ Object-Level Edit covers adding, removing, replacing, or changing individual objects within an image. ➤ Text or Symbol Edits covers changes to words, numbers, logos, and symbols within an existing image.