GPT Image 2.6 API and Pricing, Explained Through a Real Pipeline
GPT Image 2.6 is OpenAI's newest image model, released as two tiers: Sunburst for detail and Flare for speed. Both list at the same per-token rate, so the price of an image is set by the quality level, from about 2.5 cents at low to about 21 cents at max. Here is the rate card, the token counts we measured, how long each tier took, what changed against GPT Image 2.5, and how to reach either tier through a published API endpoint.
GPT Image 2.6 is OpenAI's newest model for making and editing images, and it comes as two tiers with the same inputs. Sunburst is tuned for detail: dense compositions and text-heavy layouts. Flare is tuned for speed. Both reached OpenRouter on November 4, 2026, as openai/gpt-image-2.6-sunburst and openai/gpt-image-2.6-flare. We probed them the next day through the same path the Image Generation node uses, then added both to the node and to a model page. Flare also runs in chat.
The rate card is written per image token, and the number of tokens depends on the quality level you pick. This post covers that rate card, the token counts we measured at each level, how long each tier took, what moved against GPT Image 2.5, and how to call either tier as a published API endpoint.
The spec sheet
| Capability | Sunburst | Flare |
|---|---|---|
| Input | Text and image | Text and image |
| Quality levels | Low, medium, high, extra high, max | Low, medium, high, extra high, max |
| Aspect ratios | 8, plus auto | 8, plus auto |
| Reference images | Up to 10 per node | Up to 10 per node |
| Transparent background | Yes | Yes |
| Where it runs on Treza | Canvas | Canvas and chat |
The eight ratios are 1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16 and 21:9. There is no resolution setting. Every probe at 1:1 came back as a 1024 pixel square, at every quality level, so quality here buys detail inside the frame rather than a larger frame. The transparent background option was not part of these probes.
Because the two tiers share every setting, a pipeline built on one switches to the other by changing the model on the node.
What GPT Image 2.6 costs per image
These are the listed prices as of November 2026, and they are the same on both tiers: $0.00003 per output image token and $0.000008 per input image token. That is also the rate GPT Image 2.5 lists.
The rate only becomes a price once you know how many tokens an image uses. We measured that on November 5, 2026, at 1:1, with one prompt at all five quality levels on both tiers:
| Quality | Output image tokens | At the listed rate |
|---|---|---|
| Low | 824 | about $0.025 |
| Medium | 1,449 | about $0.043 |
| High | 1,756 | about $0.053 |
| Extra high | 3,122 | about $0.094 |
| Max | 7,024 | about $0.211 |
The token counts were identical on Sunburst and Flare, apart from a two token difference at high, and each bill matched the rate card once the prompt's few text tokens were added. So the tier you pick does not change the price. The quality level does, by a factor of about eight and a half from low to max.
For scale: a hundred low quality drafts list at about $2.47. The same hundred at max list at about $21.
One note on the default. With quality left unset, our one-line prompt rendered at the low level's 824 tokens. On GPT Image 2.5 we saw the default climb with denser prompts, so set the quality yourself when you want a predictable cost.
Why low and medium cost more than on GPT Image 2.5
The per-token rate did not change between 2.5 and 2.6. The token counts at the bottom of the ladder did. Against what we measured on GPT Image 2.5 at the same 1:1 frame in September:
| Quality | GPT Image 2.5 tokens | GPT Image 2.6 tokens |
|---|---|---|
| Low | 196 | 824 |
| Medium | 439 | 1,449 |
| High | 1,756 | 1,756 |
| Extra high | 3,122 | 3,122 |
| Max | 7,024 | 7,024 |
At the listed rate, a low quality image goes from under a cent on 2.5 to about 2.5 cents on 2.6, roughly four times as much. Medium goes from about 1.3 cents to about 4.3, roughly three times. High, extra high and max cost the same on both.
If your pipeline leans on hundreds of low quality drafts, 2.5 Flare is still the cheaper draft model, and GPT Image 2 and the 2.5 tiers stay in the node's model menu. If you render at high or above, 2.6 costs you nothing extra.
Sunburst vs Flare: same price, different clock
Since the tiers bill the same, the choice between them is about time and detail. Wall time for the same prompt on November 5:
| Quality | Flare | Sunburst |
|---|---|---|
| Low | about 6 seconds | about 17 seconds |
| Medium | about 10 seconds | about 27 seconds |
| High | about 10 seconds | about 50 seconds |
| Extra high | about 23 seconds | about 82 seconds |
| Max | about 41 seconds | about 166 seconds |
One render per level on one day is not a benchmark, but the gap was consistent: Sunburst took between two and a half and five times as long. That is why Flare runs in chat and Sunburst stays on the canvas.
The prompt asked for a ceramic cup of coffee on a sunlit wooden cafe table with steam rising. At low, Flare returned a speckled stoneware cup of latte art beside a jar of white flowers, and Sunburst the same kind of cup filled with black coffee. Both are clean photographs.
At max, both tiers added a cafe chalkboard in the background that the prompt never mentioned, with handwriting you can read. Flare's carried a three line greeting. Sunburst's was a menu listing Espresso, Americano, Latte, Cappuccino and Flat White, every word spelled correctly. Legible text is the model's known strength, and it showed up unprompted.
Reference images and the edit path
With a prompt alone either tier generates. With a prompt and reference images it edits or composes, on the same node. Wire the images into the Image Generation node's reference input, or drag them onto it, and write the change as the prompt.
Our edit probe took a product still of a folded navy linen shirt and asked for the same shirt on a white marble counter in a bright studio, at low quality. Both tiers kept the shirt: the collar, the weave and the fold all carried over, set on veined marble under hard window light. Flare finished in about 10 seconds and Sunburst in about 21.
The reference billed 1,523 input image tokens, which at the listed rate is about 1.2 cents. So a low quality edit with one reference lists at about 3.7 cents. Add references where the subject has to survive, and leave them off loose concept drafts.
Flare to draft, Sunburst to finish
- Draft on Flare at low. About 2.5 cents and six seconds a frame. Ask for it by name in chat, or pick it on the node.
- Finish on Sunburst when a frame earns it. Rerun the chosen prompt at high or max, once. The prompt, the ratio and the references carry across unchanged.
- Use Flare at high when time matters more. High on Flare took about 10 seconds against 50 on Sunburst for the same price.
Where it sits in the price band
GPT Image 2.6 against the per-image prices in our earlier pricing posts (as of October 2026, at 1K where the model has a resolution setting):
| Model | Per image |
|---|---|
| GPT Image 2.6, high | about $0.053 |
| GPT Image 2.6, medium | about $0.043 |
| MAI-Image-2.6 | $0.039 |
| Nano Banana 2.1 | $0.036 |
| GPT Image 2.6, low | about $0.025 |
| MAI-Image-2.6 Flash | $0.0195 |
| Seedream 5.0 Flash | $0.018 |
| FLUX 3 Image Klein | $0.009 |
GPT Image 2.6 at low sits between the flash tier and the mid band, and at high it sits above MAI-Image-2.6 and Nano Banana 2.1. It is not the cheapest way to fill a storyboard: Seedream 5.0 Flash and FLUX 3 Image Klein do that for less. Reach for it when the frame carries words, a menu, a label or a poster, or when an edit has to keep a product exactly as it is.
What you actually pay on Treza
GPT Image 2.6 through a Treza pipeline is metered from a prepaid credit balance. One credit equals $0.01, packs start at $5, credits never expire, and there is no subscription. Treza quotes the price before a run, each image is charged for the quality it rendered at, and a failed generation charges nothing, including a prompt the provider refuses.
Calling GPT Image 2.6 as an API
Direct access means OpenAI credentials, token-level billing to reconcile, and retry logic before you have built anything. Through OpenRouter, both tiers run on the images endpoint only.
Through a Treza pipeline, GPT Image 2.6 is a node on a canvas. Place the Image Generation node, pick Sunburst or Flare, set the quality and aspect ratio, wire references into its reference input, and publish the graph as a single versioned endpoint. Your app calls the typed /invoke route with JSON in and JSON out, or points an existing OpenAI SDK at the OpenAI-compatible route. Swap the node from Flare to Sunburst later and the endpoint your product calls stays the same.
The same pipeline is reachable from an agent. Treza's hosted MCP server exposes pipelines as tools that Claude and other MCP clients call directly, so an agent can generate the frame and hand back the URL as one step of a larger task.
Putting it in a pipeline instead of a single call
A single image is rarely the deliverable. The more useful pattern is GPT Image 2.6 as the first node of several: compose the key frame on Sunburst, wire it into the First frame port of a video node, then put narration, captions, or publishing behind the video. A fuller AI video pipeline puts a script step ahead of the image and reruns the whole graph on a schedule or from an API call.
The AI image generator page has the image side set up, the AI video generator page puts a video model behind a single generate action, and the AI thumbnail generator page covers the frames where legible text matters most. For more, see how to build an AI image generation pipeline and MAI-Image-2.6 pricing.
Frequently Asked Questions
How much does GPT Image 2.6 cost per image?
The listed rate as of November 2026 is $0.00003 per output image token on both tiers. At 1:1 we measured 824 tokens at low, 1,449 at medium, 1,756 at high, 3,122 at extra high and 7,024 at max, which works out to about $0.025, $0.043, $0.053, $0.094 and $0.211 an image.
What is the difference between GPT Image 2.6 Sunburst and Flare?
Sunburst is the detail tier and Flare is the speed tier. They take the same inputs and settings and list at the same rate. In our tests Flare finished a low quality image in about 6 seconds and a max quality one in about 41, where Sunburst took about 17 and 166.
Is GPT Image 2.6 more expensive than GPT Image 2.5?
At low and medium, yes. The per-token rate is the same, but 2.6 rendered 824 tokens at low and 1,449 at medium where we measured 196 and 439 on 2.5, so those levels cost roughly four and three times as much. High, extra high and max use the same token counts and cost the same.
Do reference images cost extra with GPT Image 2.6?
They bill as input image tokens at $0.000008 each. One reference in our edit probe used 1,523 tokens, about 1.2 cents, so a low quality edit with one reference lists at about 3.7 cents. Up to 10 references fit on one Image Generation node.
Can GPT Image 2.6 edit an image I already have?
Yes. Attach the image as a reference and write the change as the prompt. In our probe both tiers kept a folded navy shirt intact, collar, weave and fold, and moved it onto a white marble counter in a bright studio.
Can I use GPT Image 2.6 in chat?
Flare, yes. Ask for GPT Image 2.6 Flare by name in Treza chat and the image comes back in the conversation. Sunburst runs on the canvas, where its longer renders at the top quality levels do not hold up a conversation.
