OpenAI Unveils 'ChatGPT Image 2.5' Powered by 'Sketch'... Sweeps No. 1 in Arena Leaderboard

1. The '@Sketch' Revolution: Beyond Text-Centric Prompts

For years, creators have wrestled with the inherent ambiguity of text-to-image prompts. Describing precise physical proportions, furniture layouts, camera perspectives, or exact clothing seams using adjectives alone frequently resulted in frustrating prompt roulette.

With ChatGPT Image 2.5, OpenAI bridges the gap between text and spatial intent. By typing @Sketch directly into the conversation box, users summon an interactive in-chat drawing canvas. You can sketch a rough room blueprint, trace out furniture placement, outline character postures, or map architectural horizons, then prompt: "Turn this into a photorealistic mid-century modern loft at golden hour."

Crucially, users do not need artistic drafting ability. The model's spatial reasoning engine translates crude outlines and compositional doodles into balanced, high-resolution compositions while honoring the user's intended layout down to the pixel.

2. Surgical Image Editing & Reference Preservation

Prior diffusion models suffered from severe "editing drift"—modifying a minor detail like changing a shirt color often destroyed facial identity, warped background architecture, or altered lighting geometry.

ChatGPT Image 2.5 radically enhances multi-turn conversational editing. Users can point, lasso, or specify exact sub-regions of an image to modify. You can alter a model's jacket or swap out a coffee mug on a desk while the rest of the composition, lighting, and textures remain locked in place.

Furthermore, reference preservation has reached production standards: when an input photo of a person or physical product is supplied, ChatGPT Image 2.5 seamlessly inserts them into fresh backgrounds and styles while preserving facial identity, skin subsurface scattering, specular reflections, and brand contours without uncanny artifacts.

3. 50% Speed Reduction & Dual Architectures: Flare vs. Sunburst

Latency has been halved compared to the previous Image 2.0 baseline. To serve both high-velocity commercial applications and fine-art studios, OpenAI launched two distinct API models:

4. Sweeping the Arena Text-to-Image Leaderboards

The performance gains translated immediately to public benchmark dominance. In the independent LMSYS Arena Text-to-Image and Image Edit evaluations, OpenAI achieved a decisive sweep:

• Rank #1: GPT-Image-2.5 Sunburst
• Rank #2: GPT-Image-2.5 Flare

Human preference raters overwhelmingly favored Image 2.5's photorealism, typography rendering, and strict adherence to negative constraints over competitive closed and open-weight diffusion systems.

5. OpenAI's Official Image Prompting Checklist

Alongside the model release, OpenAI published an official prompt engineering playbook to help creators achieve reproducible outputs without brute-force generation:

For complex scenes, OpenAI recommends specifying the scene, subject, details, and constraints separately. When including text, specify the exact phrase in quotation marks, along with its placement and font appearance.

When editing, it is best to change only one thing at a time and explicitly list everything the model should preserve. If using multiple references, you should explain the role of each image beforehand.

GPT Image 2.5 now features two models: Flare, optimized for speed, and Sunburst, optimized for maximum quality. Custom resolutions up to 4K and multiple quality levels—ranging from "low" to "max"—are also available.

6. C2PA Provenance & 3 Billion Weekly Milestone

OpenAI confirmed that over 3 billion images are now generated every week across ChatGPT and enterprise API tiers. To maintain provenance and authenticity standards, ChatGPT Image 2.5 embeds cryptographic C2PA metadata and imperceptible digital watermarks to reliably verify AI origin without visual degradation.

ChatGPT Image 2.5 is rolled out across desktop, mobile app, and web interfaces for ChatGPT Plus, Team, Enterprise, and Codex developer tiers.