GPT-Image-2.5 API
GPT-Image-2.5 (ChatGPT Images 2.5) is OpenAI's image generation and editing model family released on September 8, 2026. It features improved image detail, natural lighting, higher reference fidelity, faster generation speeds, and a split API design featuring Flare for speed and Sunburst for precision.
GPT-Image-2.5 Flare API
openai/gpt-image-2.5-flare
OpenAI’s fast default image model for generation and editing, offering sharper visuals, better text rendering, transparent backgrounds, and lower latency.
GPT-Image-2.5 API Background
Overview
GPT-Image-2.5 (product side name ChatGPT Images 2.5) is OpenAI's image generation and editing model released on September 8, 2026. Across ChatGPT and API endpoints, the GPT Image family generates over 3 billion images per week.
Development History
Released on September 8, 2026, GPT-Image-2.5 succeeds the April 2026 Images 2.0 / gpt-image-2 model. It represents the first time OpenAI has split the GPT Image API into two separate models (Flare and Sunburst) to separate speed and precision instead of relying on a single quality parameter.
Key Innovations
- Separated API offerings into gpt-image-2.5-flare (fast) and gpt-image-2.5-sunburst (high-precision)
- Added @Sketch feature in ChatGPT to generate finished renders directly from canvas hand-sketches
- Introduced preset layout Templates for posters, flyers, merchandise, and product graphics
- Added pin-comment annotations on images for targeted area editing
- Supported prompt sharing alongside image sharing in ChatGPT
- Introduced xhigh and max image quality tiers
- Improved transparent background generation and Chinese character text rendering clarity
GPT-Image-2.5 API Technical Specifications
Architecture
Available via API as two distinct model variants (gpt-image-2.5-flare and gpt-image-2.5-sunburst) through Images API endpoints (/v1/images/generations, /v1/images/edits) or the Responses API image generation tool.
Parameters
Configurable parameters include background (auto, opaque, transparent), output_format (png, jpeg, webp with compression settings), quality (auto, low, medium, high, xhigh, max), and custom dimensions.
Capabilities
- Accepts text and image inputs to output images without billing output text tokens
- Supports up to approximately 16 reference images during editing (up to ~50MB per PNG, JPEG, or WebP file)
- Supports mask editing requiring an alpha channel matching the size and format of the image
- Supports custom resolutions with side lengths in multiples of 16, maximum side <= 3840, aspect ratios up to 3:1, and pixel counts between 655,360 and 8,294,400
- Generates transparent backgrounds (PNG or WebP) for stickers, UI icons, and game assets
- Embeds C2PA metadata and invisible watermarks for provenance tracking
Limitations
- Resolutions above 2K (e.g. exceeding 2560x1440) are noted as experimental and less stable
- Individual reference images are restricted to roughly 50MB
- Mask edits require exact dimension matching and alpha channel presence
GPT-Image-2.5 API Performance
Strengths
- Up to 50% lower official latency compared to Images 2.0, with Flare reaching 2-4x speed of GPT-Image-2 in third-party testing
- Sharper visual detail, natural lighting, and richer material textures
- Higher fidelity retention of human features, facial traits, products, and key details from reference photos
- Precise localized editing that alters target areas while preserving overall composition, lighting, and subjects
- Robust multi-turn consistency that prevents quality degradation across continuous edit cycles
- Substantially improved Chinese text rendering without garbled characters on posters, packaging, and infographics
- Top rankings on LM Arena text-to-image and image editing leaderboards
Real-world Effectiveness
On LM Arena community benchmarks post-launch, Sunburst reached #1 (~1421 ELO) and Flare reached #2 (~1399 ELO) across both text-to-image and editing leaderboards. Third-party testing recorded large batch API generation times dropping from around 104 seconds down to 35–40 seconds.
GPT-Image-2.5 API When to Use
Scenarios
- Use gpt-image-2.5-flare for rapid exploration, social media images, product previews, visual search, and high-throughput batch generation
- Use gpt-image-2.5-sunburst for final advertising stills, polished product photography, strict multi-turn subject locking, and high-precision editing
- Use ChatGPT Images 2.5 for conversational editing, quick template starting points, sketch conversions, and pin-comment annotations
- Use transparent background modes for creating stickers, game assets, brand graphics, and web illustrations
Best Practices
- Default to Flare for initial drafting and throughput, then switch to Sunburst for production render execution
- In edit prompts, explicitly separate elements to modify from elements that must remain locked
- Execute multi-turn edits incrementally by changing one detail per iteration
- Keep custom output dimensions aligned to multiples of 16, within 655,360–8,294,400 pixels, and at or below 2K for optimal stability