GPT Image 2.5 pricing: what each quality tier actually costs

Published 9 September 2026 · Measured on the OpenAI API

Cost per generated image

OpenAI publishes no per-image table for GPT Image 2.5, and its model pages warn that the GPT Image 2 calculator does not estimate 2.5's token consumption. So we measured it: every figure below comes from calling the API at 1536×1024 and reading usage.output_tokens on the response, billed at $30 per million.

Both variants — gpt-image-2.5-sunburst and gpt-image-2.5-flare — consume exactly the same tokens at every tier, so they cost the same. Choose between them on quality and speed, never on budget: OpenAI describes sunburst as higher quality than GPT Image 2 and flare as comparable, and flare returned images 2 to 4 times faster in our runs.

The line item nobody counts: reference images

If you edit rather than generate from scratch — passing earlier images so a character stays the same — those images are billed as input tokens. A 1536×1024 reference costs 1,536 tokens, or $0.0123, identical on both models. That is not a rounding error. On the low tier, three references make up roughly 87% of the total cost of the call.

GPT Image 2.5 also rejects the input_fidelity parameter, so there is no setting that makes references cheaper. The only lever is sending fewer of them, which is a real trade-off when you rely on them to hold a character across scenes.

One more measured oddity: at low, a 1024×1024 square costs 196 output tokens against 158 for either 1536×1024 or 1024×1536. Square is the expensive format, not the cheap one.

What the same money actually buys

Prices only matter next to output. Below, GPT Image 2 · medium and GPT Image 2.5 · high are compared at exactly the same cost — both burn 1,372 output tokens, $0.0412 an image. Same prompt, same character, ten visual styles. The second scene of each pair was generated with that model's own first scene passed as a reference image, which is how you test whether a model holds a subject across shots.

Each style shows four images. The top row is scene 1 from both models, the bottom row is scene 2, each generated from the scene 1 directly above it. So read across to compare the two models, and read down to see what a model keeps from its own previous shot. The two models start from different seeds, so their compositions differ by design — that is not the thing being measured. Click any image to open it full size.

GPT Image 2 — scene 1

GPT Image 2.5 — scene 1

GPT Image 2 — scene 2, from its scene 1

GPT Image 2.5 — scene 2, from its scene 1

GPT Image 2 — scene 1

GPT Image 2.5 — scene 1

GPT Image 2 — scene 2, from its scene 1

GPT Image 2.5 — scene 2, from its scene 1

GPT Image 2 — scene 1

GPT Image 2.5 — scene 1

GPT Image 2 — scene 2, from its scene 1

GPT Image 2.5 — scene 2, from its scene 1

GPT Image 2 — scene 1

GPT Image 2.5 — scene 1

GPT Image 2 — scene 2, from its scene 1

GPT Image 2.5 — scene 2, from its scene 1

GPT Image 2 — scene 1

GPT Image 2.5 — scene 1

GPT Image 2 — scene 2, from its scene 1

GPT Image 2.5 — scene 2, from its scene 1

GPT Image 2 — scene 1

GPT Image 2.5 — scene 1

GPT Image 2 — scene 2, from its scene 1

GPT Image 2.5 — scene 2, from its scene 1

GPT Image 2 — scene 1

GPT Image 2.5 — scene 1

GPT Image 2 — scene 2, from its scene 1

GPT Image 2.5 — scene 2, from its scene 1

GPT Image 2 — scene 1

GPT Image 2.5 — scene 1

GPT Image 2 — scene 2, from its scene 1

GPT Image 2.5 — scene 2, from its scene 1

GPT Image 2 — scene 1

GPT Image 2.5 — scene 1

GPT Image 2 — scene 2, from its scene 1

GPT Image 2.5 — scene 2, from its scene 1

GPT Image 2 — scene 1

GPT Image 2.5 — scene 1

GPT Image 2 — scene 2, from its scene 1

GPT Image 2.5 — scene 2, from its scene 1

What we found

Character consistency is not where the gain is. GPT Image 2 already held the coat, the hair and the boots from one scene to the next across every style we tried. If you were hoping 2.5 fixes character drift at this quality level, it was not the problem to begin with.

The visible gain is scene continuity. GPT Image 2.5 carries the setting forward, not just the subject: the waterfall and autumn foliage in the photorealistic pair, the mountain peak and lake in the Pixar-style pair, the village and stone house in the Ghibli pair all survive into the second scene. For a video made of twenty shots, that is the difference between a sequence of images of the same character and a place that holds together.

And one place it is worse. On paper cutout, 2.5's extra detail pulls the image toward a painted illustration and away from the flat, layered cut-paper look that GPT Image 2 renders more honestly. We checked whether the pattern repeated on crayon illustration — it did not, 2.5 stays faithful there — so this is specific to that style rather than a general rule. Worth knowing if a flat craft aesthetic is the point of your video.

Methodology

Every number on this page was produced by calling the OpenAI Images API directly on 9 September 2026 and reading the usage object returned with each response, not by reading documentation. Generations were 1536×1024 with gpt-image-2.5-sunburst and gpt-image-2; edits passed one PNG reference of the same dimensions. The comparison images are unretouched API output, resized for the web and otherwise untouched. Token consumption can change without notice when a provider updates a model, so treat these as a snapshot and re-measure before committing a budget.

We build TubeTube, which turns written words into finished multi-scene videos, and we ran this comparison to decide which image model to put in front of our own users. GPT Image 2.5 is one of twelve image models available there, alongside Gemini 3 Pro, Flux 2 Pro and Seedream 5 Lite. If holding one character across a long video is your problem, that gets its own page: AI character consistency.

Frequently asked questions