What it is
GPT Image 2.5 is OpenAI's image generation and editing family, released on September 8, 2026. It ships as two API models: gpt-image-2.5-flare, the default for most work, and gpt-image-2.5-sunburst, which takes longer per image in exchange for tighter control across edits. The same update powers ChatGPT Images 2.5 for all ChatGPT, ChatGPT Work and Codex users.
Both models accept text and image input, add 'xhigh' and 'max' quality levels, allow custom resolutions within set limits, and gain an 'opaque' background option. In ChatGPT the release also added a Sketch tool, templates and comment-based edits.
Our review
OpenAI holds the top two places on Artificial Analysis's text-to-image arena: Sunburst at max quality scores 1222 and Flare 1207, ahead of GPT Image 2 (1186) and Google's Nano Banana 2.1 (1176). The margins are about ±10, so the two OpenAI models' ranges overlap, and Flare only just clears its predecessor.
Token pricing did not change; the quality ladder did. Rates match GPT Image 2 ($8 per million image input tokens, $30 per million output), but quality now runs low, medium, high, xhigh and max, and developers on OpenAI's forum report that max is the old high. At max, Artificial Analysis prices a 1024x1024 image at about $0.21, against $0.034 for Nano Banana 2.1 and $0.039 for MAI-Image-2.6. OpenAI's GPT Image 2 calculator does not estimate 2.5's token use, so budgets need a test batch.
Sunburst is the editing model. OpenAI pitches it for precision across chains of edits at the cost of longer generation, and one forum regular likens it to a rebranded input-fidelity setting from gpt-image-1.5. Both models sit on the Images generation and edit endpoints plus Batch. Their model pages describe use through the Responses image tool while the endpoint tables mark Responses unsupported, a contradiction to test before building on it.
Throughput is the hidden limit. API accounts on the Build tier get 20 images a minute, so a 10,000-image catalogue job takes more than eight hours; Grow allows 250. ChatGPT users face a different problem: forum posters report Codex output tagged as version 2.0 and no reliable way to tell which model a given image came from.
Where it earns its place
A brand agency producing campaign key visuals
Take one hero image through rounds of client edits
Use Sunburst on the edits endpoint so detail survives each round. At about $0.21 per max-quality square image, twenty revisions cost a few dollars, trivial next to design hours; expect slower generation than Flare.
An e-commerce marketplace with 50,000 SKUs
Generate a lifestyle shot for every listing
At max quality that is about $10,500, against roughly $1,700 on Nano Banana 2.1. Batch halves output tokens to $15 per million and lower quality steps cost less. A Build-tier account at 20 images a minute needs about 42 hours of generation.
A SaaS team adding image creation to its product
Ship in-app generation and editing for paying users
Call gpt-image-2.5-flare on the Images API and pin the snapshot gpt-image-2.5-flare-2026-09-08. API images are not used for training and abuse logs are kept 30 days unless the organisation is approved for zero data retention; uploads flagged as possible CSAM are kept for review either way.
Before you commit
- Setup
- API models gpt-image-2.5-flare and gpt-image-2.5-sunburst, with 2026-09-08 snapshots, on v1/images/generations, v1/images/edits and Batch. Quality is low, medium, high, xhigh, max or auto.
- Integrations
- OpenAI SDKs and the Images API; the same models power ChatGPT Images 2.5 in ChatGPT, ChatGPT Work and Codex.
- Data and privacy
- OpenAI does not train on API images by default and keeps abuse-monitoring logs for 30 days. Zero data retention covers both 2.5 models for approved organisations, but uploads flagged as possible CSAM are retained for manual review.
- Cost at scale
- At max quality 1,000 square images cost about $211 by Artificial Analysis's count, so 10,000 a month is about $2,100 against about $340 on Nano Banana 2.1. Batch halves image output to $15 per million tokens.
- Limits
- 20 images a minute on Build, 150 on Launch and 250 on Grow. Constraints for custom resolutions are not published on the model pages. Outputs that breach content rules are blocked after generation with a moderation error.
What's good
- First and second on Artificial Analysis's text-to-image arena (1222 and 1207)
- Sunburst holds detail across multi-step edits
- Five quality levels, custom resolutions and an opaque background option
- Zero data retention eligible on the API; not trained on by default
What's not
- About $0.21 per max-quality square image, six times Nano Banana 2.1
- Only 20 images a minute on the entry Build tier
- OpenAI's cost calculator does not cover 2.5 token use
- No weights, no self-hosting
Best for
Agencies and product teams whose images go in front of customers and justify a premium per image, especially for iterative edits.
Alternatives
GPT Image 2.5 (Flare and Sunburst): common questions
- Is GPT Image 2.5 (Flare and Sunburst) worth it?
- Sunburst and Flare lead the blind-vote rankings, at roughly six times the per-image price of Nano Banana 2.1. We rate it 4.7 out of 5.
- How much does GPT Image 2.5 (Flare and Sunburst) cost?
- API: $5 per 1M text input tokens, $8 per 1M image input and $30 per 1M image output for both Flare and Sunburst (Batch output $15). Artificial Analysis puts a 1024x1024 max-quality image at about $211 per 1,000. Included in ChatGPT plans.
- What are the downsides of GPT Image 2.5 (Flare and Sunburst)?
- About $0.21 per max-quality square image, six times Nano Banana 2.1; only 20 images a minute on the entry Build tier; openAI's cost calculator does not cover 2.5 token use; no weights, no self-hosting.
- Who is GPT Image 2.5 (Flare and Sunburst) best for?
- Agencies and product teams whose images go in front of customers and justify a premium per image, especially for iterative edits.
- How much does GPT Image 2.5 (Flare and Sunburst) cost for a team or in production?
- At max quality 1,000 square images cost about $211 by Artificial Analysis's count, so 10,000 a month is about $2,100 against about $340 on Nano Banana 2.1. Batch halves image output to $15 per million tokens.
- Is GPT Image 2.5 (Flare and Sunburst) safe for company data?
- OpenAI does not train on API images by default and keeps abuse-monitoring logs for 30 days. Zero data retention covers both 2.5 models for approved organisations, but uploads flagged as possible CSAM are retained for manual review.
- What are the best GPT Image 2.5 (Flare and Sunburst) alternatives?
- The closest alternatives we have reviewed are Nano Banana 2.1, Nano Banana Pro (Gemini 3 Pro Image), MAI-Image-2.6 and Midjourney V8.
Answers checked Oct 22, 2026.
Make GPT Image 2.5 (Flare and Sunburst)? Show this rating on your site
