GPT Image 2.5 API Pricing: Where to Run Flare and Sunburst

GPT Image 2.5 costs $30 per million image output tokens, and OpenAI does not publish a per-image price anywhere. This page converts the token rate into what one image actually costs — $0.0059 at low quality, $0.2107 at max — and lists every place you can run Flare and Sunburst today.

A 1024×1024 image costs $0.0059 (low) to $0.2107 (max) on GPT Image 2.5.
Flare and Sunburst are billed at identical token rates — the choice between them is speed
versus editing precision, not price.
max quality costs exactly what high used to cost on GPT Image 2: same money, two rungs higher.
The OpenAI free tier cannot call these models at all: rate limits start at Tier 1 — so
without a funded API account you need a host that already has one, such as
Kavel, which runs both variants on credits.

What OpenAI shipped: GPT Image 2.5 Flare and Sunburst

GPT Image 2.5 arrived as two API models, not one. The launch was covered on openai.com and in
the developer thread on community.openai.com; the version of record for this page is the API
documentation, which is what the figures below are read from. Both carry a dated snapshot id —
gpt-image-2.5-flare-2026-09-08 and gpt-image-2.5-sunburst-2026-09-08 — which is the date
stamped on the snapshot in the model reference, read 2026-09-09. OpenAI does not state
elsewhere in the API docs that this is the announcement date, so treat it as the snapshot date
it is.

The Flare model reference
and the Sunburst model reference
carry identical pricing blocks. Both add xhigh and max, which earlier GPT Image models never had — those topped out at
high. Both default to auto. Both take text and image input and return images only; neither
supports streaming, function calling, structured outputs, or fine-tuning (read 2026-09-09).

What one GPT Image 2.5 image actually costs

OpenAI publishes a token rate, not a price per image, and the docs explicitly warn that the
old calculator does not carry over: "The GPT Image 2 calculator does not estimate GPT Image
2.5 token consumption" — stated in the
image generation guide
(read 2026-09-09).

The numbers below come from OpenAI's own token calculator. The values it uses are not printed
on the page — the calculator ships as a client-side component, so the table lives in its
JavaScript bundle. Read out of that bundle on 2026-09-09, the per-quality base is
{low:16, medium:24, high:48, xhigh:64, max:96} for GPT Image 2.5 and
{low:16, medium:48, high:96} for GPT Image 2, and output tokens work out to
ceil(u × h × (2,000,000 + width × height) / 4,000,000), billed at $30 per million.

Figures derived 2026-09-09 from OpenAI's published token rates and calculator logic. They
match the one third-party per-image figure in circulation ($0.053 for 2.5 high at 1024×1024),
which is a useful check that the arithmetic is right.

Landscape and portrait cost less than square

A 1536×1024 image holds 50% more pixels than a 1024×1024 one and costs 22% less —
$0.0412 against $0.0527 at high quality. That is not a rounding artifact. Token count is
driven by a tile grid whose base number is divided by the aspect ratio, so stretching an image
away from square shrinks one side of the grid faster than the pixel count grows.

If your output is destined for a 16:9 slot anyway, generating at 1536×1024 rather than
generating square and cropping saves a fifth of the bill.

max is the new high, at the old price

GPT Image 2 high quality at 1024×1024 costs $0.2107. GPT Image 2.5 max at the same size
costs $0.2107 — the same number, because the base token count is 96 in both cases. Two rungs
of quality arrived without a price change at the top of the ladder.

Lower down the change is bigger. At medium and high the base halves (48→24, 96→48), and
because the base is squared into a tile grid, the cost drops fourfold:

low is the exception: its base is 16 in both models, so a low-quality image costs precisely
what it did before.

One caveat worth stating plainly

OpenAI's calculator uses a single token table for GPT Image 2.5 and does not split Flare from
Sunburst. The prose in the same documentation says the opposite is possible: "The models can
use different token counts for the same quality setting and share the same price per image
output token" (read 2026-09-09). So treat the table above as OpenAI's own estimate, which is
what it is, and read usage off your first real response before you budget a large run.

Is GPT Image 2.5 included with ChatGPT Plus?

Not in the sense most people mean. The two model ids documented here —
gpt-image-2.5-flare and gpt-image-2.5-sunburst — are API models, priced per token in the
API reference, with their own usage tiers. What a ChatGPT subscription includes on the
consumer side is set out on OpenAI's own plan pages and is not stated in the API
documentation, so we are not going to assert a mapping between the two here.

What the API documentation does settle, in one line, is the free question (read 2026-09-09):

The free tier is listed as not supported for these models, so a funded API account is the
documented route — or a host that has already done that for you, which is what the
image models on Kavel are.

Where you can run GPT Image 2.5

Every row below was fetched and read on the date shown, and every cell says what was actually
on the page — not what a host is assumed to support. Two entries publish a price you can act
on; the rest name the model without saying what one image costs, and inventing a number for
them would be worse than an empty cell.

The last three rows are the reason this table was worth building. All three are cited by answer
engines on this topic, and none of them names GPT Image 2.5 anywhere on the page we could read
on 2026-09-10. Being cited by an answer engine is not the same as hosting the model — worth
knowing if you are working from an AI-generated shortlist rather than checking the pages.

OpenRouter lists the model (its page names Sunburst 15 times and quotes the same $5.00 text-input
rate) but routes to providers rather than publishing its own per-image price, so it is context
rather than a place to compare cost. apidog.com, eesel.ai and vercel.com cover request testing
and deployment, not access.

Running GPT Image 2.5 without an API key

Every route above that needs a key needs a funded OpenAI account first, because the free tier
is not supported. The alternative is a host that has already paid that bill and resells access
by the image.

Kavel is the second row in that table: both variants wired, a flat
25 / 40 / 60 credits per image at 1K / 2K / 4K, the same rate whether you generate or edit, and
no API key or usage tier in the way. A new account starts with 40 credits — one 2K image, or one
1K image with change. Rates are on the pricing page.

The trade is the obvious one: direct against OpenAI gets you raw token billing and your own
rate limits, a host gets you access without an account at that host's markup. Which is cheaper
depends entirely on volume — and the break-even is lower than people expect, because
GPT Image 2 at the same quality still costs $0.2107 against 2.5's $0.0527,
so the model you pick moves the bill four times more than the host you pick. Both are on the
image side at the same per-image credit rate, and editing costs what generation costs
— an image-to-image run is not billed differently.

The cheapest way to run GPT Image 2.5

If you are generating at volume and already have a funded OpenAI account, going direct is the
floor: $0.0059 an image at low quality, $0.0132 at medium. Nothing resells below cost.

If you are generating occasionally, the account is the expensive part, not the image. A Tier 1
account caps at 5 images a minute and has to be funded before it will run anything at all, so
for a few dozen images a month the per-image credit route costs less in total than standing up
and topping up an API account.

And if you only need one or two, generate at medium rather than high. The quality gap is
one rung; the price gap is 4×.

Migrating from the GPT Image 2 API to GPT Image 2.5

The request shape is unchanged. Both models sit on the same endpoints — v1/images/generations,
v1/images/edits, and the image generation tool inside v1/responses — so the migration is
usually a string swap from gpt-image-2 to gpt-image-2.5-flare or gpt-image-2.5-sunburst.

Four things to re-check after the swap:

• Your quality constant. xhigh and max did not exist before. Code that clamps quality
  to high now silently caps you two rungs below the ceiling.
• Your cost model. If anything in your stack estimates spend from GPT Image 2's token
  table, it is now wrong in both directions: 4× too high at medium and high, and blind to the
  two new rungs.
• Size divisibility. Width and height must both be divisible by 16, the pixel budget caps
  at 8,294,400, and the long edge caps at 3,840.
• Streaming costs. Each partial image emitted through partial_images adds 100 image
  output tokens on top of the final image.

Pin the dated snapshot ids (gpt-image-2.5-flare-2026-09-08) if you need reproducible output
across a release; use the undated ids if you would rather follow improvements. If you are
moving video work at the same time, the video models are priced on a separate scale
— per second rather than per image — so the arithmetic here does not carry across.

If you would rather not maintain the integration at all, the same two models are already wired
on the AI image maker, billed per image instead of per token. The
showcase gallery has real output from each model currently wired, and
Nano Banana 2 is the cheaper alternative if text rendering is not what
you need.

What we could not verify about GPT Image 2.5 pricing

Billing on failed, refused, and moderated requests is not stated in the model reference or the
image generation guide, and we are not going to infer it. If a request is rejected by
moderation, whether the image output tokens are still billed is a question for OpenAI support
or your own dashboard after a test call.

Per-image pricing at fal.ai, Morphic, Felo, wavespeed, modellix, renoise and cellcog is
likewise unstated in page source as of 2026-09-09. The OpenAI announcement post itself
(openai.com/index/introducing-chatgpt-images-2-5/) returned HTTP 403 to us, so every OpenAI
fact here comes from the API documentation rather than the launch blog.

FAQ

Is GPT Image 2.5 Flare cheaper than Sunburst?

No. OpenAI bills both at $30 per million image output tokens, $8 per million image input
tokens, and $5 per million text input tokens (read 2026-09-09). The documentation notes that
actual token consumption can differ between models at the same quality setting, so measure
usage on a real call before assuming the per-image cost is identical too.

How much does one GPT Image 2.5 image cost?

At 1024×1024: $0.0059 at low, $0.0132 at medium, $0.0527 at high, $0.0937 at xhigh, and
$0.2107 at max. Landscape and portrait at 1536×1024 or 1024×1536 cost about 22% less than
square at the same quality.

Can I use GPT Image 2.5 without an OpenAI API key?

Yes, through a host that supplies its own. Kavel runs both Flare and Sunburst on credits with
no key and no usage tier. Directly against OpenAI a key is mandatory, and the free tier is
listed as "Not supported" for these models.

Is GPT Image 2.5 included with ChatGPT Plus?

A ChatGPT subscription covers image generation in ChatGPT. The API models are billed
separately against API credit; the two balances are not shared.

Can I use GPT Image 2.5 output commercially?

Commercial rights come from the terms you accept with whoever you run the model through, not
from the model itself, and they are not stated in the API model reference. Read OpenAI's usage
terms if you are calling the API directly, or the host's terms if you are not. On Kavel the
pricing page sets out what a paid plan includes.

Am I charged for a generation that fails or gets refused?

Not stated. Neither the model reference nor the image generation guide says whether image
output tokens are billed on a request that errors or is blocked by moderation, and we are not
going to guess. Test one call and read the usage on your own dashboard before you plan around
it.

Do GPT Image 2.5 images come out watermarked?

The word "watermark" does not appear anywhere in the model reference or the image generation
guide (checked 2026-09-09), so OpenAI does not state that API output carries one. What a
reseller hands you is a separate question — some hosts return a marked file on free tiers and
an unmarked one on paid. Check the host's own terms rather than assuming either way.

Do credits expire?

On the API side you are buying credit against an OpenAI account, and expiry is set by OpenAI's
billing terms, not by the image model. On Kavel, credit granted by subscription and credit
bought outright are tracked separately and have different lifetimes — the current rules are on
the pricing page.

What payment options are there, and can I cancel?

Against the API there is nothing to cancel: OpenAI bills per token for what you actually
spend. On Kavel a one-off credit pack is an alternative to subscribing, so you can pay once
without a recurring charge to cancel later. Both payment routes, and the refund policy that
applies to each, are on the pricing page.

What replaced the gpt-image-2 API?

Nothing was removed — gpt-image-2 still works. GPT Image 2.5 Flare and Sunburst are the
recommended models for new integrations, and they add the xhigh and max quality settings
that GPT Image 2 does not have.

Sources

OpenAI API model reference for gpt-image-2.5-flare
and gpt-image-2.5-sunburst,
and the image generation guide,
all read 2026-09-09. Token-count logic read from the calculator component served on that guide
page. Cross-checked against published per-image figures at apidog.com, eesel.ai and
openrouter.ai. Host rows fetched and read 2026-09-10 — fal.ai, Morphic, Felo, gpt-img.com, wavespeed.ai, and
the three that turned out not to name the model (modellix.ai, renoise.ai, cellcog.ai). Launch
coverage checked but not used for figures: 9to5mac.com, unite.ai, xenospectrum.com,
intelligentliving.co, m.36kr.com, eu.36kr.com, wallstreetcn.com, community.openai.com,
vercel-docs.vercel.sh, synapse.garden — all repeat OpenAI's headline numbers without adding a
per-image figure. Kavel credit rates read from the live
site configuration 2026-09-10.

Last updated: September 2026