What GPT Image 2.5 actually costs
OpenAI publishes token rates, not a price per picture. Converting one into the other is the whole job — here are the rates and the arithmetic.
The most useful thing to understand about GPT Image 2.5 pricing is that there is no price per image. OpenAI meters tokens — text in, image in, image out — and an image is billed by how many output tokens it consumed, which varies with both the requested size and the quality tier. That design is fine for a bill and awkward for a budget, because it means you cannot quote a client a per-asset cost until you have measured your own workload.
The good news for anyone already running the previous model: the rates did not move. OpenAI's model pages list identical token rates for GPT Image 2.5 and GPT Image 2, which makes this a rare upgrade with a strictly better speed-per-dollar figure and no migration cost to justify.
Published token rates
| Token type | Rate | Cached | Unit |
|---|---|---|---|
| Text input | $5.00 | $1.25 | per 1M tokens |
| Image input | $8.00 | $2.00 | per 1M tokens |
| Image output | $30.00 | — | per 1M tokens |
Both Flare and Sunburst carry these same rates. Text output is not billed, because these models emit images rather than text. The cached image input rate is the one people forget: at $2.00 against $8.00 per million, re-sending the same reference photo across an edit chain is four times cheaper than treating each turn as a cold call.
Working estimate: cost per image
Since the token rates match GPT Image 2 exactly, the measured per-image figures from that model are the best available proxy for GPT Image 2.5 at the three established tiers.
| Quality | 1024×1024 | Portrait / landscape | Typical use |
|---|---|---|---|
| low | ~$0.006 | ~$0.006 | Drafts, thumbnails, A/B sweeps |
| medium | ~$0.053 | ~$0.041 | Client-facing everyday work |
| high | ~$0.211 | ~$0.165 | Marketing and hero assets |
| xhigh | not published | not published | New in this generation |
| max | not published | not published | New in this generation |
Four monthly bills, worked through
Solo creator
~$3200 images / month · Drafts at low, 40 finals at medium
Well inside a starter credit grant. Cost is not the constraint at this volume; your time is.
Content team
~$463,000 images / month · 2,400 low drafts, 600 medium finals
The draft-then-promote discipline is what keeps this two figures instead of three.
Same team, no discipline
~$6333,000 images / month · Everything at high
Identical output volume, 13x the bill. This is the single most expensive habit on the platform.
Product integration
~$30050,000 images / month · Low tier, batched
Batch plus the low tier is what makes user-facing generation viable as a free feature.
The third row is the point of the table. Two teams generating the same three thousand images can differ by a factor of thirteen on the invoice purely through tier discipline, and nobody notices until finance asks a question. Draft cheap, promote deliberately.
Five levers that actually move the bill
Draft at low. The spread between the low and high tiers on the previous generation was roughly thirty-five times per image. If you are going to discard the first six attempts, discard cheap ones. This single habit dominates every other optimisation.
Cache your image inputs. Cached image input bills at $2.00 per million tokens against $8.00 cold. On an editing workflow that re-sends the same reference across a dozen turns, that is most of your input cost gone.
Batch anything non-interactive. The Batch endpoint is supported for both GPT Image 2.5 models, and it is the standard route to reduced rates for work that does not need a synchronous response. Overnight catalogue generation should never run interactively.
Prefer portrait and landscape where composition allows. On the previous generation, non-square renders came in cheaper at medium and high — around $0.041 and $0.165 against $0.053 and $0.211. The mechanism is token count, so the saving should carry.
Route by workload, not by prestige. Flare and Sunburst cost the same per token. Sending single-shot work to Sunburst buys latency, not quality you can see, and on a large queue that latency has its own cost in infrastructure and patience.
Third-party gateways
A cluster of resellers front this model family with flat per-image pricing, which solves the predictability problem at the cost of a margin. Some of them advertise resolution tiers the official API does not list, reachability from regions OpenAI does not serve, or access without an OpenAI account. Those are real advantages for some teams.
Two cautions. First, flat pricing is only cheaper if your mix skews to expensive tiers; at the low tier the official API is usually the cheaper option outright. Second, gateways advertising "reverse-engineered" access to a consumer channel are not operating under a supported contract, and the thing they are selling can disappear without notice. For anything a business depends on, the official API or a named cloud partner is the defensible choice.