GPT Image 2.5 Is Live on Runable: Flare and Sunburst
Key Takeaways
- OpenAI shipped GPT Image 2.5 on September 8, 2026, in two API models:
gpt-image-2.5-flarefor speed andgpt-image-2.5-sunburstfor maximum quality and editing precision. - Both models are available on Runable right now. Ask for an image in any task and the agent generates, edits, and drops the result straight into the deck, site, or document it is building.
- Headline gains are up to 50% lower latency, two new quality tiers (
xhighandmax), native 4K output up to 3840x2160, and noticeably better multi-turn editing and subject preservation. - Token rates are unchanged from GPT Image 2: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens on both variants.
OpenAI has been shipping fast this month. GPT-6 Astra landed on September 3. Five days later, on September 8, 2026, the company released GPT Image 2.5 across ChatGPT and the API, and split its image stack into two models for the first time: Flare for everyday speed, Sunburst for the jobs where quality decides everything.
Both are live in Runable. This post covers what actually changed, which of the two you should reach for, what it costs, and how to put it to work inside a real project instead of a chat window.
What OpenAI shipped on September 8
GPT Image 2.5 replaces a single general-purpose image model with a two-model lineup, both snapshotted 2026-09-08 .
GPT Image 2.5 Flare is the small model, optimized for speed. OpenAI describes its image quality as comparable to GPT Image 2, delivered faster. Flare is the default choice for high-volume work: social variants, thumbnails, product shots at scale, draft passes before a final render.
GPT Image 2.5 Sunburst is the base model, optimized for quality. It produces sharper detail, more natural lighting, and greater control than GPT Image 2, and OpenAI positions it specifically for workflows where editing precision matters most. If GPT Image 2 was falling short on a complex brief, Sunburst is the model that fixes it.
Both models accept text and image inputs, both output images only, and both support the full quality ladder: low , medium , high , xhigh , max , and auto . The xhigh and max tiers are new in this generation. Both also support transparent backgrounds natively, which matters more than it sounds: a logo or product cutout that comes out of the model with a clean alpha channel saves a whole compositing step.
Resolution moved too. GPT Image 2.5 takes auto or a custom WIDTHxHEIGHT , with common presets at 1024x1024, 1536x1024, 2048x1152, and 3840x2160 for 4K landscape. Custom sizes have to keep each edge under 3,840 pixels, keep both edges as multiples of 16, stay inside a 3:1 aspect ratio, and land between 655,360 and 8,294,400 total pixels. OpenAI flags anything above 3,686,400 pixels as experimental, so 4K is real but not yet boring.
Flare or Sunburst: which one should you pick?
Start from what your current workflow already does, not from the model names.
| Your situation | Start with | Why |
|---|---|---|
| GPT Image 2 already meets your quality bar | Flare | Same quality class, lower latency, cheaper per accepted image in practice |
| GPT Image 2 falls short on a complex brief | Sunburst | Higher ceiling on detail, lighting, and control |
| Precise multi-step editing of one asset | Sunburst | Built for editing precision and subject preservation |
| High-volume generation, tight deadlines | Flare | Speed is the whole point of the small model |
OpenAI's own guidance is a two-pass loop. Establish quality first with Sunburst, then test Flare against the same prompts and inputs. If Flare clears your bar and improves latency, switch. If it does not, keep Sunburst. The company is explicit that a speed improvement on one workload does not transfer to another, so measure on your actual prompts rather than trusting a benchmark screenshot.
One practical warning from the docs that applies to both models: repeated edits can still drift details you meant to freeze. Restate your constraints on every pass, and when a region has to stay pixel-identical, composite the approved edit back into the original instead of asking the model to preserve it.
What actually got better
Three things changed in a way you will notice within about ten minutes of using it.
Latency dropped, by up to 50% depending on the workload. That number comes from OpenAI's release material and it is workload-dependent, but the direction is consistent. For anyone generating a dozen variants of the same asset, the compounding effect is larger than the headline.
Editing got more faithful. Both models improved on precise editing and subject preservation, which is the failure mode that made earlier image models frustrating for real production work: you ask for one change and the model quietly redraws the face, shifts the product geometry, or resets the lighting. GPT Image 2.5 holds more of the frame steady while it changes the part you named.
Noise artifacts went down. The visible grain and smear that showed up in earlier GPT image outputs, particularly in flat regions and gradients, is reduced. That is the difference between an asset you can ship and an asset you have to clean up.
What GPT Image 2.5 costs
Pricing is identical across both variants, and identical to GPT Image 2's token rates.
| Metric | Price per 1M tokens |
|---|---|
| Text input | $5.00 |
| Cached text input | $1.25 |
| Image input | $8.00 |
| Cached image input | $2.00 |
| Image output | $30.00 |
Text output is not billed, because these models emit images rather than text. Worth noting: OpenAI says the GPT Image 2 token calculator does not estimate GPT Image 2.5 consumption, so if you are budgeting from a spreadsheet built on the old model, rebuild it from a real sample of your own requests.
The important cost question is not price per million tokens anyway. It is cost per accepted image. A faster, cheaper model that needs four retries to clear review is more expensive than a slower one that lands on the second attempt. Track the retry rate alongside the invoice.
How to use GPT Image 2.5 on Runable
On Runable you do not pick an endpoint or wire a request. You describe the image you want as part of the task you are already doing, and the agent handles model selection, generation, and placement.
Start a task and say what you need. "Build me a landing page for a cold brew brand and generate the hero shot" is a complete instruction. The agent writes the page, generates the image with GPT Image 2.5, and puts it in the hero slot at the right dimensions. The same applies inside a deck, a report, a carousel, or a product listing.
Iterate in the same conversation. Because GPT Image 2.5 is built for multi-turn editing, follow-up instructions work the way you would expect: "same shot, warmer light, move the can to the left third" changes those things and leaves the rest alone. Say what must stay fixed and the model holds it.
Ask for the format you need. Transparent PNG for a logo, 4K landscape for a print-bound hero, a tight square for social. State it in plain language and the agent maps it to the right size and background setting.
The reason this is worth doing inside an agent rather than a chat window is that images are almost never the deliverable. The deck is the deliverable. The site is the deliverable. Generating an image in one tool, downloading it, and dragging it into another is the slow part of the job, and it is the part Runable removes. Runable has a free tier to try it, with Pro at $20/mo as the most popular plan and Max at $100/mo for heavy usage.
Prompting GPT Image 2.5: five rules that carry
OpenAI published a prompting guide alongside the release. The parts that matter most in practice:
Name the result and its use. "Product photograph for an e-commerce listing" gets you further than a description of the object alone, because it tells the model the conventions to follow.
Describe visible detail, not mood. Materials, lighting direction, color, and medium beat adjectives. Ask for "photorealistic" or "real photograph" explicitly when that is the goal, because the model will not assume it.
Quote exact text and say how many times it appears. Put required wording in quotes, describe its position and typography, and then ask for no additional text. Check the spelling in the output every time.
Separate the change from the constraints. For edits, say "change only X" and list what must survive: identity, geometry, layout, labels, lighting. Naming the exclusions is what stops the drift.
Assign roles to references. When you pass multiple inputs, identify each one by number and purpose, subject or style or background, and explain how they combine.
GPT Image 2.5 versus Nano Banana Pro
Google's Nano Banana Pro is the obvious alternative, and it is also available on Runable, so this is a choice you can make per task instead of per subscription.
GPT Image 2.5 Sunburst is the stronger pick for photorealism, precise text rendering, transparent-background assets, and edits where identity and product geometry have to survive intact. It rewards structured prompts with labeled sections.
Nano Banana Pro holds character and product identity more consistently across many separate, independent generations, fuses more reference images in a single call, and rewards natural descriptive sentences over rigid fields. If you are producing a long series of images that all have to feature the same character, that consistency is the deciding factor.
The honest summary: reach for Sunburst when one image has to be right, and for Nano Banana Pro when forty images have to match. Flare covers the volume in between.
If you want to see how GPT Image 2.5 fits into a broader creative workflow, the best AI image generators comparison is a useful reference for how this generation of models stacks up. For teams already using Runable, the complete 2026 guide to Runable walks through how to set up tasks that combine image generation with documents, decks, and sites. And if GPT-6 Astra's release earlier this month changed how you think about model selection, the GPT-6 Astra launch post covers what that model adds and when to reach for it over the image stack. For small businesses deciding where AI image generation fits in a wider toolset, AI tools for small business ranks the options by job type.
FAQ
What is GPT Image 2.5?
GPT Image 2.5 is OpenAI's image generation and editing model family, released September 8, 2026. It ships as two API models, gpt-image-2.5-flare for speed and gpt-image-2.5-sunburst for quality, and is also available inside ChatGPT and on Runable.
What is the difference between Flare and Sunburst?
Flare is the small model optimized for speed, with image quality comparable to GPT Image 2. Sunburst is the base model optimized for quality, with sharper detail, more natural lighting, and greater control. Both support the same quality settings, sizes, and transparent backgrounds.
Can GPT Image 2.5 generate 4K images?
Yes. It supports custom resolutions up to 3840x2160, with each edge capped at 3,840 pixels and both edges as multiples of 16. OpenAI marks outputs above 3,686,400 total pixels as experimental, so verify results at that size before committing to it in production.
How much does GPT Image 2.5 cost?
Both models are billed at $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens, with cached input at $1.25 and $2.00 respectively. Text output is not billed. On Runable, image generation is covered by your plan credits instead.
Should I migrate from GPT Image 2?
If GPT Image 2 already meets your quality bar, test Flare first and look for the latency gain. If it does not, test Sunburst and confirm the quality before optimizing anything else. Keep prompts, references, and dimensions unchanged for the first comparison so you are measuring the model and not your own edits.
Related Reading
- Best AI Image Generators in 2026: A Complete Comparison
- GPT-6 Astra Is Here: OpenAI's New Frontier Model, Now in Runable's Ultra Mode
- How to Use Runable: The Complete 2026 Guide
- AI Tools for Small Business: 10 Best in 2026 (Tested)
- Introducing Runable 2.0: The Best Way to Work with AI
- Claude Fable 5.1 Is Out: Anthropic's Frontier Model, Now in Runable's Ultra Mode
Stop generating images in one tool and pasting them into another
Describe what you are building, and Runable generates the images with GPT Image 2.5 and puts them where they belong: the hero slot, the slide, the product page, the carousel. Start free, and upgrade to Pro at $20/mo when the output becomes part of your week.
