OpenAI Ships ChatGPT Images 2.5 with Flare and Sunburst

OpenAI released ChatGPT Images 2.5 on September 8, 2026, replacing the Images 2.0 generation that shipped in April with a model the company describes as sharper, faster, and markedly better at editing an image without disturbing the parts you did not ask it to change. Two API variants launched alongside it — gpt-image-2.5-flare and gpt-image-2.5-sunburst — and OpenAI reports that generation latency has fallen by up to 50% versus Images 2.0. Independent Arena rankings placed the two models first and second across all three image leaderboards within a day of launch.

Intermediate

What Changed in the Model

OpenAI frames Images 2.5 around three improvements rather than a single headline capability. The first is reference fidelity: given a photo of a person, place, or product, the model is better at carrying distinctive features through into new settings, styles, and compositions. The second is precision editing — changing one element while leaving the subject, composition, and surrounding treatment intact. The third is multi-turn consistency, so that a fifth edit in a long conversation still respects the first four and does not visibly degrade the image.

The model also handles transparent backgrounds and more complex layouts, and OpenAI says images containing real-world information are more accurate. Scale gives the changes some weight: the company reports that more than 3 billion images a week are now created across ChatGPT Images and the GPT-Image API models.

Inside ChatGPT, the release adds Sketch — invoked by typing @Sketch — which lets users draw a rough layout directly in the conversation and use it as a visual guide. Templates cover common formats such as posters, flyers, merchandise, and product photos. Users can now place comments directly on an image to scope an edit, and share the prompt behind an image so others can rerun it with their own inputs. Images 2.5 is rolling out to all ChatGPT, ChatGPT Work, and Codex users across every tier on desktop, mobile, and web.

Two Models in the API

The developer story is a split rather than a single upgrade. GPT-Image-2.5 Flare is the default: OpenAI positions it as delivering higher-quality output than GPT-Image-2 at half the latency, aimed at social and creator content, visual search, prototyping, and high-volume generation. GPT-Image-2.5 Sunburst is the slower, more precise option, intended for production campaign creative and polished product imagery where control across successive edits matters more than throughput.

Both accept text and image input, emit images only, and support six quality settings (low, medium, high, xhigh, max, auto). They are reachable through v1/images/generations, v1/images/edits, and as the image-generation tool in the Responses API. Neither supports streaming, function calling, structured outputs, or fine-tuning. Dated snapshots are pinned as gpt-image-2.5-flare-2026-09-08 and gpt-image-2.5-sunburst-2026-09-08.

Pricing is identical for the two models, and unchanged from GPT-Image-2:

Token type	Input (per 1M)	Cached input (per 1M)	Output (per 1M)
Text	$5.00	$1.25	not billed
Image	$8.00	$2.00	$30.00

OpenAI notes that while the token rates match GPT-Image-2, the GPT-Image-2 cost calculator does not estimate 2.5 token consumption — so per-image cost has to be measured, not extrapolated. Neither model is available on the free tier; rate limits run from 100,000 tokens and 5 images per minute at Tier 1 up to 8,000,000 tokens and 250 images per minute at Tier 5.

Adobe, Manus, Runway, and Higgsfield AI were named as early customers. Axultan Alimkulov, Head of Product at Higgsfield AI, said what impressed the team most about Flare was how well it understands what not to change.

Provenance and Safety

The system card reports an unsafe generation rate of 1.09% for Sunburst and 1.41% for Flare under automated adversarial testing, against a 1.64% baseline for Images 2.0. The safety stack has four layers: LLM-based policy checks that refuse a request before it reaches the generator; a multimodal safety reasoning model that screens both text and image inputs; the same monitor checking the finished image before it is shown; and continuous offline and online monitoring, with separate evaluation stacks for higher-risk categories involving minors.

On provenance, OpenAI continues to attach C2PA metadata under the C2PA Conformance Program, and has added Google DeepMind’s SynthID invisible watermarking across ChatGPT, Codex, and the API. SynthID survives the kinds of transformation — screenshotting, recompression, cropping — that strip C2PA metadata, so the two mechanisms cover different failure modes.

What This Means

The competitive picture moved quickly. Arena reported Sunburst at #1 and Flare at #2 across the Text-to-Image, Image Edit, and Multi-Image Edit leaderboards, with both models improving on GPT-Image-2 (medium) in all three. The gains are largest where the release is aimed: Sunburst picked up +40 points in Text-to-Image, +59 in single-image edit, and +81 in multi-image edit, with Flare at +18, +30, and +47 respectively. Artificial Analysis had not yet published an independent score for either model as of September 9, where GPT-Image-2 (high) still held the top text-to-image slot at Elo 1178.

For anyone building on these models, the more consequential detail is the split itself. Until now the choice within a GPT-Image generation was mostly a quality dial; Flare and Sunburst are two models at the same token price with different latency and precision profiles, which turns model selection into a routing decision made per workflow rather than per account. A high-volume thumbnail pipeline and a campaign-asset pipeline now have genuinely different right answers, and because pricing is identical, the trade-off is purely wall-clock time against edit precision.

The SynthID adoption is worth noting separately. Invisible watermarking developed at Google DeepMind now runs across OpenAI’s image surfaces — a rare instance of two competing labs converging on shared provenance infrastructure rather than parallel proprietary schemes. For institutions that need to determine whether an image was machine-generated, a common detection substrate is more useful than two incompatible ones.

Related Coverage

• OpenAI Launches ChatGPT Images 2.0 with 2K Output and Reasoning Mode — the April 2026 predecessor that Images 2.5 replaces
• Google Launches Nano Banana 2: Pro-Quality Image Generation at Flash Speed — the Gemini 3.1 Flash Image release that held the Arena runner-up spot
• Z-Image: Alibaba’s Efficient 6B Open-Source Image Generation Model — the open-weight end of the same market
• OpenAI Launches GPT-6 Astra, Its First ‘Critical’ Cyber Model — OpenAI’s other September release

This post was drafted with AI assistance and reviewed by RITS staff.

Sources

• Introducing ChatGPT Images 2.5 — OpenAI, September 8, 2026
• GPT-Image-2.5 Flare model reference — OpenAI API documentation
• GPT-Image-2.5 Sunburst model reference — OpenAI API documentation
• ChatGPT Images 2.5 System Card — OpenAI Deployment Safety Hub
• Introducing ChatGPT Images 2.5 — Simon Willison, September 8, 2026
• OpenAI Releases ChatGPT Images 2.5 With Sketch and Two New API Models — Unite.AI
• Arena.ai leaderboard results for GPT-Image-2.5 — Arena, September 8, 2026