ChatGPT Images 2.5 suits teams doing iterative, comment-based photo edits and multilingual text-in-image work, generating up to 4K output up to 50% faster than the prior Images 2.0 release. It replaces DALL-E across every OpenAI product and adds a Sketch reference-drawing mode for people who prefer sketching over prompting.
ChatGPT Images 2.5 is OpenAI's autoregressive image generation model, released September 8, 2026, leading the Artificial Analysis Text-to-Image Arena at 1187 Elo. It ships as two API snapshots, Flare for speed and Sunburst for precision editing, replacing the diffusion-based DALL-E line entirely.
Provider: OpenAI · Family: GPT Image
Input modalities: text, image · Output: image
About ChatGPT Images 2.5
ChatGPT Images 2.5 is OpenAI's image generation model, released September 8, 2026, the newest entry in the GPT Image line that began with GPT Image 1 (the "GPT-4o image generation" launch of March 25, 2025). Unlike the diffusion-based DALL-E models it replaced, GPT Image runs on an autoregressive transformer that generates an image token by token. OpenAI ships 2.5 as two API snapshots, gpt-image-2.5-flare and gpt-image-2.5-sunburst (both dated 2026-09-08); ChatGPT calls the pair "Images 2.5" and routes each prompt to whichever fits its complexity (Flare for speed, Sunburst for precision). Parameter count is undisclosed; OpenAI publishes no token-based context window here as it does for GPT-5 or o3. OpenAI has not published SWE-bench, MMLU or GPQA scores for an image model, so the usual text-model benchmark set does not apply. The number that does exist comes from Artificial Analysis's Text-to-Image Arena, a blind human-preference vote: as of September 2026, GPT-Image-2.5 Flare (max quality) tops the leaderboard, ahead of Sunburst (max) at 1179 Elo and the prior GPT Image 2 (high) at 1171. The system card reports a 1.09% unsafe-generation rate for Sunburst and 1.41% for Flare on its adversarial test set, sexual content blocked 89.1% of the time on Sunburst, extremism 0% on both. Output tops out at 4K across nine aspect ratios and five quality tiers (low/medium/high/xhigh/max) plus auto. Input is text and image (edits, reference-driven generation); output is image only, so no text response is billed. The headline new feature is Sketch: typing "@Sketch" lets a user draw inside ChatGPT and feed that drawing in as a visual reference. Comment-based editing flags one region of a generated image for a targeted change instead of a full prompt rewrite, and a shared-prompt option hands another user the exact instructions and photos to rerun an idea. Inpainting is supported on both snapshots via /v1/images/edits. Both snapshots share one price sheet: text input $5 per 1M tokens, reference-image input $8 per 1M tokens (both with a cached-input discount), and every generated image billed as image-output tokens at a separate, higher rate. OpenAI publishes no fixed token count per image, so cost varies by resolution and quality tier rather than a flat per-image price; neither gets a batch discount. ChatGPT bundles usage into its Free, Plus, Pro, Business and Enterprise subscription tiers instead of billing per image. The API sits at /v1/images/generations and /v1/images/edits with usage-tier rate limits; Azure OpenAI Service also hosts the line. An upstream policy check screens a prompt before generation and a separate safety model screens the output before display; a blocked request returns a refusal, not an image. Every image carries both C2PA metadata and Google DeepMind's SynthID watermark. OpenAI's API platform holds a SOC 2 Type II report (Jul 2025-Jun 2026) plus ISO 27001/27701 certification. Teams needing legible multilingual text inside the image, Latin and CJK scripts alike, see more consistent results here than from Midjourney, still rated weaker on in-image text. Iterative, comment-based edits that must leave earlier elements untouched are Sunburst's target case. Teams wanting a hand-illustrated style with minimal prompting, or a bigger free tier, are better served by Midjourney or Google's Nano Banana Pro; teams needing open weights to self-host should use Black Forest Labs' FLUX.2, since GPT Image ships closed-weight and API-only. Images 2.5 is the fifth named GPT Image release, after GPT Image 1 (March 2025), GPT Image 1-mini (October 2025, ~80% cheaper), GPT Image 1.5 (December 2025, ~4x faster, ChatGPT's default), and GPT Image 2 (April 2026, reasoning-driven prompt adherence). It adds Sketch, comment-based edits, shared prompts, the xhigh/max tiers, and up to 50% lower latency than Images 2.0. OpenAI retired DALL-E alongside this transition: dall-e-2/3 stopped serving May 12, 2026, and the standalone DALL-E GPT was retired in late August 2026.
Pricing
Billed by token category, identical for both gpt-image-2.5-flare and gpt-image-2.5-sunburst: text prompt input $5/1M tokens ($1.25/1M cached), reference-image input $8/1M tokens ($2/1M cached), generated image output $30/1M tokens (text output is not billed). OpenAI does not publish a fixed token count per image, so no flat per-image price exists; there is no batch-API discount.
Key Features
- Sketch Reference Drawing: Typing @Sketch lets a user draw directly inside ChatGPT and use that sketch as a visual reference for the generated image, instead of describing the composition in words.
- Comment-Based Editing: Drop a comment on one region of a generated image to request a targeted change there, while the rest of the image stays untouched across the edit.
- Five Selectable Quality Tiers: Choose low, medium, high, xhigh, or max quality per request (plus an auto mode), trading generation cost and latency against fidelity up to 4K resolution.
- Dual API Snapshots (Flare/Sunburst): Flare targets everyday speed while Sunburst auto-escalates for complex edits inside ChatGPT; both are billed at the same per-token rate and reachable independently through the API.
- Dual Provenance Watermarking: Every generated image carries both OpenAI's C2PA-conformant metadata and Google DeepMind's SynthID invisible watermark, so origin can be checked two ways.
Pros
- Tops the Artificial Analysis Text-to-Image Arena leaderboard (Flare, max quality), ahead of the prior GPT Image 2.
- Comment-based, region-specific edits and the new Sketch mode cut out full-prompt rewrites for iterative work.
- OpenAI reports faster generation and unchanged per-token pricing versus the prior release.
Cons
- The exact token cost of a single generation isn't published up front, so per-image cost can't be known until after a request runs.
- Closed weights and API-only access mean no self-hosting, fine-tuning, or air-gapped deployment; Black Forest Labs' FLUX.2 is the open-weight alternative.
- The system card's own adversarial testing still shows a 1.09-1.41% unsafe-generation rate depending on the snapshot, so a moderation layer is still worth budgeting on top of OpenAI's.
Benchmarks
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What are ChatGPT Images 2.5's pricing plans in 2026?
OpenAI meters cost by token rather than by image: $5 per 1M text-prompt tokens ($1.25 cached), $8 per 1M reference-image tokens ($2 cached), and $30 per 1M generated image-output tokens, identical for both the Flare and Sunburst snapshots. Because no fixed token count per image is published, the same prompt can cost a different amount depending on resolution and the chosen quality tier. Inside ChatGPT the feature is bundled into the Free, Plus, Pro, Business and Enterprise subscription price rather than billed per image.
Does ChatGPT Images 2.5 have a free plan?
Free ChatGPT accounts get a small number of image generations per rolling 24-hour window (historically 2 to 3 under the prior default), while Plus, Pro, Business and Enterprise accounts get progressively higher or effectively unlimited generation subject to OpenAI's abuse guardrails. The API itself has no free tier: billing starts from the first request at the per-token rates in the pricing FAQ above.
Which tools compete with ChatGPT Images 2.5 in 2026?
Google's Nano Banana Pro is the closest rival on overall image quality and ships a small free daily allotment inside the Gemini app, which the OpenAI API does not offer. Black Forest Labs' FLUX.2 is the strongest open-weight alternative for teams that need to self-host or fine-tune rather than call a closed API. Midjourney remains a pick for a distinctive hand-illustrated look with minimal prompting, though it is still rated weaker on rendering legible in-image text.
What separates ChatGPT Images 2.5 from Nano Banana Pro?
Both currently sit at the top of the Artificial Analysis Text-to-Image Arena, with GPT-Image-2.5's two snapshots ahead of the prior GPT Image 2. Independent reviewers rate Nano Banana Pro ahead on raw scene realism and factual grounding (correct product names, accurate on-image units), while the GPT Image line is rated ahead on multilingual typography, rendering clean Latin and CJK text inside the image itself. Neither model ships with open weights, so the choice comes down to which specific edit or text-rendering behavior fits the job.
How do you set up ChatGPT Images 2.5?
Inside ChatGPT, it needs no setup: type a prompt, or type @Sketch to draw a reference first, and the app automatically routes the request to Flare or Sunburst. Through the API, a developer needs only an OpenAI API key and a POST to /v1/images/generations or /v1/images/edits with a model name of gpt-image-2.5-flare or gpt-image-2.5-sunburst, picking a quality tier to trade cost against fidelity from the first call.
Top Alternatives
- FLUX.2: Pick FLUX.2 if you need open weights you can self-host or fine-tune; pick ChatGPT Images 2.5 if you want the higher Arena score and built-in ChatGPT access.
- FLUX 3: Pick FLUX 3 if your workflow needs native video and audio output alongside images; pick ChatGPT Images 2.5 for image-only generation with the higher Arena score.