GPT Image 2

OpenAI

GPT Image 2 (gpt-image-2) is an OpenAI image generation and editing model, released on April 21, 2026 with the default snapshot gpt-image-2-2026-04-21 and available through the Image API and the Responses API. It supports flexible resolutions up to 4K with low, medium and high quality tiers, and per-image pricing varies by resolution and quality. OpenAI's model list places GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare above it, and names them as replacements for older GPT Image models.

Full Specifications

All specsBest valueCheapestHeavy workloadMax capabilityDifferences only

	GPT Image 2OpenAI $0.006/$0.053/$0.211 per image (1024x1024, low/medium/high quality); $0.005/$0.041/$0.165 at 1024x1536 or 1536x1024 - API pricing
Identity
Developer	OpenAI
Released	Apr 2026
Status	GA
Licence	Proprietary
Self-hostable	No
Cost
Price per image	Text-to-image (Image API): $0.006/$0.053/$0.211 (low/medium/high quality) at 1024x1024; $0.005/$0.041/$0.165 at 1024x1536 or 1536x1024; additional/arbitrary sizes are billed via image+text token usage rather than a flat per-image price
Cheapest plan	No subscription plan - pay-as-you-go via the OpenAI API only, billed per output image/token (see Price per image)
Free tier	Not offered via the API - the Free usage tier lists image generation as "Not supported" in the model's own rate-limit table, and no free/trial image credits are documented; available at no cost, subject to "limited and slower image generation" rate limits, on ChatGPT's Free plan (the app's current image-generation engine)
Capacity
Max resolution	Up to 3840x2160px (4K landscape) / 2160x3840px (4K portrait) - max edge <=3840px, both edges multiples of 16px, total pixels between 655,360 and 8,294,400; outputs above 2560x1440 (2K) are "experimental"; standard sizes 1024x1024, 1536x1024, 1024x1536 also supported
Aspect ratios	Arbitrary WIDTHxHEIGHT (both edges divisible by 16, long:short ratio between 1:3 and 3:1) rather than a fixed enum; popular presets 1024x1024 (1:1), 1536x1024/1024x1536 (3:2/2:3), 2048x2048 (2K square), 2048x1152 (2K landscape), 3840x2160/2160x3840 (4K)
Images per request	Up to 10 per request via the Image API generations endpoint (n parameter, min 1, max 10); the Responses API image-generation tool returns 1 image per call by default (iterate via multi-turn conversation / previous_response_id)
Capability
Inpainting	Yes - prompt-guided masked editing (mask via input_image_mask in the Responses API tool, or the mask parameter on the Image API edits endpoint); "entirely prompt-based" - the model uses the mask as guidance but may not follow its exact shape with complete precision
Image-to-image	Yes - up to 16 reference images per request (images array parameter) for editing/reference-based generation; gpt-image-2 always processes image inputs at high fidelity (the input_fidelity parameter is not adjustable for this model, unlike earlier GPT Image models)
Text rendering	Yes - OpenAI's launch materials claim "near-perfect" text accuracy and improved multilingual rendering (incl. Japanese, Korean, Chinese, Hindi, Bengali); the API guide's own Limitations section still notes the model "can still struggle with precise text placement and clarity" on complex prompts
Style controls	Not offered as a dedicated parameter - the Images API `style` field (vivid/natural) is documented as dall-e-3-only and not supported for gpt-image-2; style transfer only via text prompting and reference images
Character consistency	Not offered as a dedicated named feature/parameter - consistency for recurring characters/faces comes from supplying reference images via image-to-image editing (up to 16); Microsoft Foundry's model listing separately highlights "Face preservation" as a key capability, while OpenAI's own Limitations section still warns the model "may occasionally struggle to maintain visual consistency for recurring characters"
Rights
Commercial rights	Yes - OpenAI's Terms of Use assign full ownership of Output to the user ("you own the Output ... We hereby assign to you all our right, title, and interest"), subject to the Usage Policies
Training-data opt-out	Yes on both surfaces, but different defaults - API usage (and ChatGPT Business/Enterprise) is excluded from training by default, opt-in only; individual ChatGPT/Codex users are opted IN by default but can opt out via Settings > Data Controls ("do not train on my content" / turn off "Improve the model for everyone")
Access
API	Yes - dedicated Image API endpoints (v1/images/generations, v1/images/edits) plus the Responses API image_generation tool; the code sample on this guide calls client.images.generate(model="gpt-image-2", ...) directly, and both endpoints are listed as supported on the model's own capability page
Chat app	Yes - powers ChatGPT's image generation as "ChatGPT Images 2.0" (launched alongside the API model), available on every ChatGPT plan including Free
Cloud marketplaces	Yes - Microsoft Foundry (Azure), listed as a "Direct from Azure" model, Generally Available; not found on AWS Bedrock or Google Cloud Vertex AI marketplaces as of this verification
Fine-tuning	Not offered - "Fine-tuning: Not supported" on the model's own capability page; no v1/fine-tuning support for gpt-image-2
Measured quality
LMArena Elo	1381 (checked Nov 2026)
LMArena rank	Rank 3