GPT Image 2.6 holds the price for a third generation — because the pricing moved off the rate card
TL;DR: On 4 November 2026 OpenAI released gpt-image-2.6-sunburst and gpt-image-2.6-flare — the third GPT Image generation to ship at exactly the same token rates: $8 per million image input tokens, $2 cached, $30 per million image output, $5 per million text input. Fourteen months of price stability on a frontier image line reads like good news, and the headline is accurate. The reason it is accurate is the story. Everything genuinely new in 2.6 is a cost dial OpenAI will not quantify: a complexity field that defaults to high, the most expensive of three settings, under a note stating it “has no separate fee”; SVG output billed by two models, one of which is never named and has no rate-card row; and animated GIF defaulting to nine frames with no statement anywhere of how frames bill. The rate card has not moved in three generations because the price is no longer kept there.
What shipped
OpenAI’s API changelog entry for 4 November reads, in full: “Released GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare for image generation and editing, with adjustable complexity, SVG generation, and animated GIF generation. Added markup-guided editing for supported GPT Image 2.5 and 2.6 models.” Both models are live on v1/images/generations and v1/images/edits.
Note what is absent: a price. There was no pricing announcement because there was no pricing change. On OpenAI’s pricing page, read 5 November, the two 2.6 rows and the two 2.5 rows carry identical figures on every line — image input $8.00, cached image input $2.00, image output $30.00, text input $5.00, cached text input $1.25 per million tokens. Both 2.5 models remain listed. Neither has a deprecation entry. This is an addition, not a swap.
The rate card has not moved in three generations
We made this point about GPT Image 2.5 in September: OpenAI held token rates flat against gpt-image-2, which meant the choice between Flare and Sunburst carried no rate-card signal at all, and the real cost lever was the quality parameter — a 35x spread from low at roughly $0.006 per 1024x1024 image to max at roughly $0.211, shipping on a default of auto. That ladder crosses Google’s flat per-image rate rather than sitting above or below it, which is why GPT Image 2.6 vs Nano Banana Pro has to fix the parameters before it can answer anything.
2.6 repeats the pattern a generation later. Three releases, one set of numbers. On most price lists that would be the end of the analysis. Here it is the beginning, because a flat rate card on a token-billed product tells you the price per token and nothing about how many tokens a default request consumes — and that is the quantity OpenAI has changed, twice, without changing a price.
complexity defaults to the most expensive setting
The new field controls “how much the image model plans before rendering”. The guide’s own framing: lower complexity “can suit focused edits, such as changing a background or a sign”, higher complexity “can help with dense text, complex layouts, and relationships between objects”.
Three levels — low, medium, high — and then the sentence that matters:
The supported complexity levels for Sunburst and Flare are
low,medium, andhigh. If you omitcomplexity, both models usehigh.
The default is the top rung. A team that upgrades to 2.6 and changes nothing else is running the most expensive planning setting on every request.
Immediately below it, OpenAI adds: “Adjusting complexity has no separate fee.” That statement is true, and it is narrower than it sounds. There is no surcharge attached to a complexity level — but there is also no published token count for any level, so the sentence describes the rate card and is silent on the invoice. Planning before rendering is work, work emits tokens, and tokens are the billing unit. OpenAI’s own caution in the preceding paragraph points the same direction: “Higher complexity can increase latency and introduce unwanted changes, so compare results for your task.” On a token-metered model, latency is rarely uncorrelated with token count.
So the honest statement of what a buyer knows is this: 2.6 added a dial that plausibly moves cost, defaulted it to maximum, and documented it with a fee disclaimer and no measurements. We cannot size the effect from the documentation, and we are not going to guess at a multiplier. What we can say is that “no separate fee” is not the same claim as “no cost difference”, and the two are easy to read as one.
SVG output is billed by two models, one of them unnamed
This is the first genuinely new billing shape in the GPT Image line, and it is three sentences long:
Markup-guided edits use the same token rates as other edits, with no extra fee. For SVG output, you pay the image model’s rate for the reference image and the authoring model’s rate for the tokens it uses to produce the SVG. The total depends on how many tokens each model uses.
Requesting a vector file triggers two billable stages: the image model renders a reference raster, then an “authoring model” writes the paths. The phrase “authoring model” appears exactly once in the entire image generation guide. It is never identified, and no model by that description has a row on the pricing page. A buyer therefore cannot compute an SVG’s cost in advance, and cannot even establish which rate card the second half of the bill is drawn from.
Two structural limits belong next to that, because they constrain what you can build: SVG output “does not support streaming, partial images”, and “the image edit endpoint does not support SVG output, and an SVG is not a supported raster input for image edits” — so vectors are generation-only in both directions. If your pipeline is iterative, SVG does not fit it yet. OpenAI’s suggested use is narrow and sensible: “graphics that need editable paths, shapes, and colors, such as icons and illustrations.”
Animated GIF defaults to nine frames
Set output_format: "gif" on POST /v1/images/generations and two new fields appear:
| Field | Range | Default |
|---|---|---|
frame_count | 2–16 frames | 9 |
frame_duration_ms | 20–2000 ms, rounded to 10 ms | 200 |
size | auto, or WIDTHxHEIGHT in multiples of 16 | auto |
A nine-frame default means a default request produces nine finished frames at whatever size you asked for, and size “describes each finished frame”. OpenAI publishes no GIF rate, no per-frame rate, and no sentence anywhere stating how frames bill. The documentation does not let you determine whether a default animation costs the same as one still image or nine of them — a range wide enough to decide whether a feature ships.
The guide does impose a budget: “The frame count, frame dimensions, and number of animations must fit the model’s size limits and the total request pixel budget.” A pixel budget is a strong hint that frames are not free, but a hint is not a rate. GIF, like SVG, is generation-only — the edit endpoint does not support it — and it drops streaming, partial images, custom output_compression and prompt_rewriter.
What to do before shipping any of this
The migration itself is optional and unpressured: both 2.5 models remain listed at the same rates with no retirement date, which is the opposite of the position developers face on the GPT-4 series leaving the API in January. The work worth doing is not the version bump.
- Pin
complexityandqualityin the same change. Inheriting both defaults means runninghighplanning under a rendering rung OpenAI selects for you. Together they are the two largest cost variables in this API and both ship permissive. - Meter SVG and GIF on a test key first. Neither has a computable price. Generate a hundred of each, read the actual token usage, and derive your own per-unit figure before either goes behind a user-facing button.
- Take markup-guided editing, which is the clean win. A transparent PNG of your annotations passed as
markuptov1/images/edits, applied to the first input image, at “the same token rates as other edits, with no extra fee” — and OpenAI backported it to supported 2.5 models, so it needs no migration. - Do not re-run your vendor comparison yet. Our ranked image generator guide and the OpenAI image review put GPT Image’s
highrung below Nano Banana Pro’s standard rate and itsmaxrung above it. 2.6 does not move those crossings, because it did not move the rates. It adds two formats whose position on that chart cannot be plotted at all.
The bottom line
A third generation at one price is a real and unusual piece of restraint, and OpenAI deserves the credit for it. But the inference most buyers will draw from it — that image generation costs what it cost in September — does not follow. The rate card has been stable precisely because the cost decisions have migrated off it and into request parameters: first quality, with its 35x range and its auto default, and now complexity, defaulted to its top setting, plus two output formats with no derivable price at all.
The pattern is now consistent enough to treat as policy. When OpenAI ships an image generation, the rate stays and the defaults get more expensive. Read the parameters, not the price list.
Token rates, parameter defaults and ranges were read from OpenAI’s pricing page, API changelog and image generation guide on 5 November 2026. Per-image cost figures for the 2.5 quality ladder are carried over from our September analysis and will vary with prompt length, reference images and output dimensions. Verify against your own usage before committing budget.
Frequently asked questions
Is GPT Image 2.6 more expensive than 2.5?
Not on the rate card, and the two generations are listed side by side at identical numbers: image input $8.00 per million tokens, cached image input $2.00, image output $30.00, text input $5.00 and cached text input $1.25 — the same figures gpt-image-2.5-sunburst and gpt-image-2.5-flare carry, and the same ones the line has had since gpt-image-2. That is the third generation in a row at one price. What can make 2.6 cost more than 2.5 on the same workload is not the rate but the request: 2.6 adds a complexity field that defaults to high, and higher complexity means the model plans more before rendering. Planning emits tokens and tokens are what you are billed for, so a request that was priced under 2.5 can cost more under 2.6 with no number on the price list having moved.
What does the complexity parameter do, and what does it cost?
It controls, in OpenAI's words, "how much the image model plans before rendering" — low for focused edits like changing a background or a sign, high for dense text, complex layouts and relationships between objects. The supported levels are low, medium and high, and the guide states plainly: "If you omit complexity, both models use high." So the default is the most expensive of the three. OpenAI also states "Adjusting complexity has no separate fee", which is true and narrower than it reads: there is no surcharge per complexity level, but there is no published token count for any level either, so the fee-free claim describes the rate card rather than the invoice. OpenAI's own warning points the same way — "Higher complexity can increase latency and introduce unwanted changes" — and latency on a token-billed model is usually a proxy for tokens. Set complexity explicitly and measure both ends on your own prompts.
How is SVG output billed?
By two models, and OpenAI names only one of them. The guide states: "For SVG output, you pay the image model's rate for the reference image and the authoring model's rate for the tokens it uses to produce the SVG. The total depends on how many tokens each model uses." The API first creates a reference image, then an authoring model writes the vector paths. That authoring model is mentioned exactly once in the entire guide, is never identified, and has no row on the pricing page — so there is no way to compute an SVG's cost in advance, or even to know which rate card the second half of the bill comes from. Two further limits are worth knowing before you build on it: SVG output does not support streaming or partial images, and the image edit endpoint does not support SVG at all, in either direction — you cannot produce an SVG from an edit, and an SVG is not a valid raster input to one.
What does an animated GIF cost?
Unknown, and the default is the part to notice. Set output_format to gif on POST /v1/images/generations and the animation fields are frame_count, accepting 2 to 16 with a default of 9, and frame_duration_ms, accepting 20 to 2000 with a default of 200. Nine frames is nine finished images at the frame size you asked for. OpenAI publishes no GIF-specific rate, no per-frame rate, and no statement of whether a nine-frame animation bills as one image output or nine, so a buyer cannot tell from the documentation whether the default costs 1x or 9x a still at the same dimensions. Until OpenAI says, treat GIF generation as unpriceable and meter it on a test key before putting it behind anything user-facing. GIF is generation-only; the edit endpoint does not support it.
Should I migrate from 2.5 to 2.6?
There is no deadline pressure, so take the time to measure. Both 2.5 models remain on the pricing page at unchanged rates and neither has a deprecation entry, so this is an addition rather than a replacement — unlike the GPT-4 series, which is leaving the API on a dated schedule. The case for moving is the new capability set: adjustable complexity, SVG, animated GIF, and markup-guided editing, which OpenAI backported to supported 2.5 models in the same release and bills at "the same token rates as other edits, with no extra fee". The case for waiting is that 2.6 carries no independent benchmark scores yet, exactly as 2.5 did not at launch, and that its two headline formats have no computable price. If you do migrate, pin complexity and quality explicitly in the same change — inheriting both defaults means running the most expensive planning setting underneath a rendering setting OpenAI chooses for you.
Sources
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.