GPT-5 Image Mini: Cheap Drafts and Bulk Image Jobs

Most image generation budgets are blown on pictures nobody keeps. A designer asks for twelve variations, picks one, and throws away eleven. A product team generates previews for every item in a catalogue and approves a fraction. For that kind of work, the cost of a single throwaway image matters far more than peak quality, and that is the job GPT-5 Image Mini is shaped for.

What the Mini pairing is

GPT-5 Image Mini is an OpenRouter listing, openai/gpt-5-image-mini, available since October 2025. Its catalogue description says it combines GPT-5 Mini for language with GPT Image 1 Mini for image generation. OpenAI released gpt-image-1-mini at DevDay on 6 October 2025, and its model page describes it as a cost-efficient version of GPT Image 1, served on the image generation, image edit and batch endpoints.

From the catalogue, the model takes text, images and files as input and returns text and images, with a 400,000-token context window, reasoning, JSON mode and structured outputs. Tool calling is not listed. In shape it mirrors GPT-5 Image; the difference is that both halves are the smaller, cheaper models.

The price gap, in OpenAI's own numbers

OpenRouter lists GPT-5 Image Mini at $2.50 per million input tokens and $2 per million output tokens at the time of writing, against $10 and $10 for the full GPT-5 Image. As with any model that mixes text and image tokens, those headline rates do not translate directly into a cost per picture.

The clearest view of the difference comes from OpenAI's per-image prices for the underlying image models, billed directly through OpenAI:

• Low quality, 1024 x 1024: $0.005 for gpt-image-1-mini against $0.011 for gpt-image-1.
• Medium quality, 1024 x 1024: $0.011 against $0.042.
• High quality, 1024 x 1024: $0.036 against $0.167.
• High quality, portrait or landscape: $0.052 against $0.25.

The ratio widens as quality rises: roughly half the price at low quality and under a quarter at high quality. A medium-quality Mini image costs about the same as a low-quality full-size one. OpenRouter's combined model will not land on exactly these figures, so measure with the usage.cost field it returns on each response, but the direction is reliable.

What you give up

A cheaper image model is cheaper for a reason. The trade-offs to look for in testing:

• Fine detail and text in images. Small lettering, dense diagrams and intricate textures are where smaller image models usually show their limits first.
• Prompt adherence on complex briefs. A request with many specific constraints is more likely to have one dropped.
• Interpretation. GPT-5 Mini is a smaller language model than GPT-5, so a vague brief or a long reference document gets a less careful reading before the image is drawn.
• Age. Both underlying models date from 2025, and OpenAI now lists GPT-Image-2.5 variants as its current image models.

None of these show up in a price table, and none can be judged from one sample. Generate the same twenty briefs on both models and compare them side by side.

A draft-then-finish workflow

The most effective way to use the Mini is not as a replacement for a stronger image model but as the first stage in front of one:

1. Explore cheaply. Generate several low or medium-quality variations with GPT-5 Image Mini to settle composition, layout and style.
2. Select. A person or an automated check picks the direction worth keeping.
3. Finish. Regenerate only the chosen concept at high quality, on the Mini if it holds up or on a stronger model if it does not, using the draft as a reference image.

Most of the images in that pipeline are drafts, so most of the spend lands at the cheapest rate. The same principle applies to text models, and routing cheap work to cheap models covers it in more depth.

Bulk jobs are the other natural fit: placeholder art, thumbnails, social previews and catalogue images where consistency and volume matter more than polish. If those jobs do not need conversational interpretation, calling gpt-image-1-mini directly through OpenAI's image endpoints, which also support batch, may be cheaper still, because you skip the language model entirely.

Mini or full: a decision rule

Pick GPT-5 Image Mini when the images are drafts, previews or high-volume assets, when a person will review them anyway, or when cost per image is the constraint. Pick the full GPT-5 Image when the image ships as it is, when it must render text or fine detail reliably, or when the brief needs the larger model to interpret a document or reference carefully.

If neither model passes your quality bar, the problem is generation rather than tier, and newer image models such as GPT-5.4 Image 2 are the comparison to run next.