Moving an image pipeline from GPT Image 2.5 to 2.6 is a model-string change, not a repricing. OpenAI released 2.6 on November 4, 2026, in the same two variants as 2.5 — Sunburst for detailed work and Flare for fast iteration — and the token rates did not move. A 1024x1024 high-quality image still costs about $0.211 in output tokens.
Three things need checking before a migration. Batch is not available on 2.6. OpenAI's materials describe the Responses image tool differently for the two versions. And AIHubMix's published gateway schema does not expose the new controls yet, so if you call through the gateway, speed is the part you can use today.
The speed claim comes from AIHubMix's model page: the Images 2.6 family generates 50% faster, without a stated baseline, and Flare can return a low-quality image in as little as six seconds. That is a vendor figure, not a benchmark, so plan around the direction rather than a stopwatch.
Before you start
Three facts about the pipeline decide the answer:
- Whether it calls OpenAI's Image API directly or AIHubMix's gateway. The new controls are documented on the OpenAI side; the gateway schema is a separate question.
- Whether it uses Batch or the Responses image tool. Batch keeps the decision on 2.5; Responses requires a test.
- Whether the work is draft-heavy. That is where speed and the $0.006 low-quality image matter.
What changed
| Area | GPT Image 2.5 | GPT Image 2.6 |
|---|---|---|
| Price | $30/M image output | $30/M image output |
| Speed | Not stated | 50% faster, vendor claim |
| Complexity control | No | low, medium, high |
| SVG output | No | Yes, generation only |
| Animated GIF output | No | Yes, generation only |
| Markup-guided edits | Yes, also supported | Yes |
| Batch API | Supported | Not supported |
| Responses image tool | Listed in the 2.5 changelog | Not supported |
| Endpoints | Images API | Images API |
The 50% figure comes from AIHubMix's model page; it is not an independent measurement. OpenAI's model pages and changelog contain no benchmark for it. Treat it as a direction, not a stopwatch.
Sunburst or Flare
The two variants share token rates, so list price does not decide between them. OpenAI positions Sunburst for detailed visual work and editing precision, and Flare for fast, everyday generation. The practical split is a two-stage workflow: draft with Flare at low quality, then render the chosen frame with Sunburst at medium or high.
The image generation guide treats complexity as separate from quality. It controls how much planning the model does before rendering, has three levels, defaults to high when omitted, and carries no separate fee. Lower complexity suits focused changes such as a background or a sign. Higher complexity can help with dense text and complex layouts, but it can also add latency and introduce changes you did not ask for. Quality still controls the rendering itself.
The migration is one string
Both models use the Images API, and the endpoints do not change. For a direct OpenAI call, the only edit is the model ID:
- "model": "gpt-image-2.5-sunburst"
+ "model": "gpt-image-2.6-sunburst"
Use gpt-image-2.6-flare where the pipeline wants Flare. On AIHubMix the same model-ID switch applies, but check the gateway section below before you rely on complexity, SVG, GIF, or markup.
Price and real cost
All five GPT Image 2.x models share the same token rates, according to the OpenAI pricing page. The differences between them are in token consumption, not list price.
| Model | Text input | Image input | Image output | Cached image input |
|---|---|---|---|---|
| GPT Image 2.6 Sunburst | $5/M | $8/M | $30/M | $2/M |
| GPT Image 2.6 Flare | $5/M | $8/M | $30/M | $2/M |
| GPT Image 2.5 Sunburst | $5/M | $8/M | $30/M | $2/M |
| GPT Image 2.5 Flare | $5/M | $8/M | $30/M | $2/M |
| GPT Image 2 | $5/M | $8/M | $30/M | $2/M |
| GPT Image 1.5 | $5/M | $8/M | $32/M | $2/M |
The cached rate needs a caveat. OpenAI says it applies only to GPT Image 2 and 2.5 images generated through the Responses API, not to direct Images API requests, including edits. The pricing page does not extend that note to 2.6.
Per-image output estimates from AIHubMix, which exclude input tokens, are identical on 2.5 and 2.6:
| Quality | 1024x1024 | 1024x1536 | 1536x1024 |
|---|---|---|---|
| low | $0.006 | $0.005 | $0.005 |
| medium | $0.053 | $0.041 | $0.041 |
| high | $0.211 | $0.165 | $0.165 |
| auto | $0.211 | $0.165 | $0.165 |
auto is an estimate: OpenAI says the cost depends on the image the model generates.
A realistic run makes the numbers concrete. Forty low-quality drafts at 1024x1024 cost 40 × $0.006 = $0.24. Four high-quality finals at the same size cost 4 × $0.211 = $0.844. The 44 images cost $1.084 in output tokens; input tokens are extra. If the finals are 1536x1024, they cost 4 × $0.165 = $0.66, and the output cost is $0.90. The same session on 2.5 costs the same, because the per-image estimates are identical.
SVG output bills differently: the image model's rate applies to the reference image, and the authoring model's rate applies to the tokens it uses to produce the SVG, so the total depends on both. Partial images during streaming add 100 image output tokens each.
The compare page puts the two generations side by side.
Failure modes
Batch is not available on 2.6
GPT Image 2.5 supports v1/batch, and Batch halves the image output rate to $15 per million tokens. GPT Image 2.6's model page lists Batch as not supported, and the Batch pricing table omits both 2.6 models. A nightly image job has no 2.6 path; keep it on 2.5.
Responses pipelines need a test
The 2.5 changelog lists the Responses API image generation tool alongside the Image API. The 2.6 changelog lists only v1/images/generations and v1/images/edits, and the 2.6 model pages mark Responses as not supported. If your code calls 2.5 through Responses, verify the 2.6 path before changing the model ID.
Gateway parameters are a schema question
AIHubMix's published schema for 2.6 accepts png, jpeg, and webp output and does not list complexity, markup, frame_count, or frame_duration_ms. It also does not list svg or gif as output formats. This is a schema gap: the features are documented for OpenAI's Image API. If you need them today, call OpenAI directly. If you need one key and one bill, check the schema again before assuming the gateway will pass them through.
Quality and size docs disagree
The image guide lists low through max as quality options for both 2.6 and 2.5, but the API reference attaches xhigh and max to 2.5 snapshots only, and AIHubMix's per-image table prices just low, medium, high, and auto. Custom dimensions are documented for 2.5, not for 2.6. The three recommended sizes are 1024x1024, 1536x1024, and 1024x1536.
What you cannot verify
- The 50% speed claim has no baseline and no independent measurement; the same applies to the six-second Flare figure. Both are vendor statements.
- Per-image estimates exclude input tokens, and
autodepends on the generated image. - Whether AIHubMix accepts the new parameters cannot be settled from the schema alone; the schema does not list them.
- Output quality is not measured here. The decision above rests on price, endpoints, and documented capabilities.
Coming from 1.5
GPT Image 1.5 retires on December 1, 2026, according to the deprecations page, and OpenAI names 2.5 as the replacement. The move is not uniformly cheaper. At 1024x1024, high quality goes from $0.133 on 1.5 (OpenAI's historical table) to $0.211 on 2.6 (AIHubMix's estimate), about 59% more; medium goes from $0.034 to $0.053. Low quality gets cheaper, from $0.009 to $0.006. Those figures are the same on 2.5, so the jump is the cost of leaving 1.5's quality tiers, not a 2.6 premium.
Switch, wait, or stay
Switch now if you generate many drafts and discard most of them: Flare at low quality is where the speed claim and the $0.006 draft price pay off. Switch if you need SVG icons or animated GIFs and call the Image API directly. If you are on 1.5, move before December 1; 2.5 is OpenAI's named replacement and carries the same per-image estimates as 2.6, so choose 2.6 only for the speed and the new output types.
Wait if you run Batch jobs, if your pipeline uses the Responses image tool, or if you depend on the new parameters through AIHubMix. None of those blocks has a published date.
Stay on 2.5 if it already works. OpenAI has announced no retirement date for 2.5, and it has markup-guided editing too. The upgrade is about speed, output types, and complexity control, not about keeping the model alive.
Trying both is one model-ID change on AIHubMix — gpt-image-2.6-sunburst or gpt-image-2.6-flare instead of the 2.5 ID — and the quick start covers the request shape.
Keep reading: the GPT Image 2.6 series
- The tutorial takes these facts into the request: which calls change, why Batch disappears, and a draft-to-final workflow with a cost line at each step. GPT Image 2.6 API tutorial: drafts, finals, and migration gotchas
Top comments (0)