20 Image Models, One Prompt: Costs Differ 22x — and One Setting Changes 35x (2026)
Quick verdict
OpenRouter ran 20 image generation models against identical prompts at 1024x1024 with default settings and published the real bills: $0.006 to $0.134 per image, a 22x spread, with OpenAI's gpt-image-2 cheapest at $0.00599 and Google's gemini-3-pro-image most expensive. The finding that matters more for budgets is that changing the quality parameter on the same model took one image from about $0.006 to $0.211 — roughly 35x — and latency from 12 seconds to 123.
The Test
OpenRouter ran 20 image generation models through its platform on September 11, 2026 with an identical prompt, identical size, and default settings, then recorded what each call actually billed. The prompt was deliberately plain: a studio packshot of a glass bottle, hard key light, no text, at 1024×1024.
The headline number is a 22x spread: $0.006 to $0.134 per image.
| Model | Cost per image |
|---|---|
| openai/gpt-image-2 | $0.006 (cheapest) |
| black-forest-labs/flux.2-klein-4b | $0.014 |
| sourceful/riverflow-v2.5-fast | $0.018 |
| Qwen image models | ~$0.03 |
| Grok image models | $0.05–0.06 |
| sourceful/riverflow-v2.5-pro | $0.064 |
| recraft/recraft-v4.1-vector (SVG) | $0.08 |
| ByteDance Seedream Pro | ~$0.09 |
| google/gemini-3-pro-image | ~$0.134 (most expensive) |
The first surprise is the cheapest entry. ChatGPT Plus costs $20 a month and OpenAI's reputation is not one of thrift — yet on the API side, gpt-image-2 came in below every competitor at just under six-tenths of a cent for a one-line prompt.
The Number That Actually Matters Is 35x
If you take one thing from this study, it should not be the 22x spread between models. It should be this: the same model, same prompt, same 1024×1024 size, moved from about $0.006 to about $0.211 when the quality parameter changed to high.
That is roughly 35x — and latency went from 12 seconds to 123.
Default and low were billed identically in this test, so the cost cliff sits specifically at the high setting. Which means the model-selection decision everyone argues about is often worth less than the configuration decision nobody checks. A team that picks the cheapest model and then passes quality: high spends more than a team that picks an expensive model on default settings.
Sort by price per call and you have answered the wrong question.
Why the Expensive Models Cost More
The cost differences are not arbitrary, and three explanations cover most of the gap.
Output size. ByteDance's Seedream models default to 2048×2048 while most others produce 1024×1024. Four times the pixels is four times the work, and the bill reflects it. If you are paying for 2048 and downscaling for social, you are buying resolution you discard.
Extra capability. gemini-3-pro-image sits at the top of the range partly because it can return text alongside the image. In the test it did so on 3 of 3 calls — the most consistent of the candidates. For pipelines that need an image and a caption, alt text, or a description in one request, that reliability has real value.
Parameter defaults. The least visible and most expensive factor, as the quality finding shows.
The Operational Data Is the Buying Guide
The pricing table gets the attention, but the integration details will decide more implementations. OpenRouter published both.
Reference image limits — how many input images a model will accept for editing:
| Limit | Models |
|---|---|
| 16 | OpenAI image models |
| 14 | Gemini 3.x image models, Seedream 5.0 |
| 10 | sourceful/riverflow-v2.5-pro |
| 8 | FLUX.2 Flex, Pro, Max |
| 4 | FLUX.2 Klein, Qwen Image 3 |
| 3 | Grok Imagine |
| 1 | Recraft, Krea, MAI Image |
If your workflow composites several product photos into one scene, a model capped at one reference is disqualified regardless of price.
Text plus image output: only 9 of 52 catalog image models support both modalities — all Gemini image models and OpenAI's gpt-5-image models, called through the chat completions endpoint. Reliability varied sharply in testing: gemini-3-pro-image returned text on 3 of 3 calls, gpt-5-image on 2 of 3, gemini-3.1-flash-image on 1 of 4, and gemini-3.1-flash-lite-image on 0 of 2.
Output format is not safe to assume. In this run FLUX.2, Grok, and Seedream returned JPEG; OpenAI, Qwen, Microsoft, and Krea returned PNG; Sourceful and Recraft returned WebP; recraft-v4.1-vector returned SVG. Only 12 of 52 catalog models accept an output_format parameter, so read the media_type on each response instead of inferring an extension.
For teams that need editable vectors — logos, icons, print assets — the SVG output from a Recraft model at $0.08 is a category of its own rather than a price point on the same curve.
The Metric You Should Actually Track
Cost per call is a supplier's unit. The number your budget cares about is cost per accepted image.
A $0.006 image that needs five attempts costs $0.03 and two minutes of someone's attention. A $0.064 image that lands first time costs $0.064 and nothing else. The cheap model still wins on that arithmetic — but only if it actually lands, and the test deliberately does not grade quality, so it cannot tell you.
The practical metric is straightforward: total image spend in a period divided by images that passed review, with human review time logged separately. Two things follow from measuring that instead of list prices:
- Retry rate dominates unit cost at the cheap end of the market.
- Draft and final work should use different settings. There is no reason for an ideation pass to run at
quality: high— that is where the 35x trap does its damage.
The Caveats
Four limits bound how far these numbers travel:
- One prompt, one size, one day. This is an observation from a single run, not a permanent price list. Model routing and pricing change.
- Prompt length changes the bill on per-token models — gpt-image-2's own range ran from $0.006 to $0.0135 depending on how much text you send.
- Default settings only. A tuned pipeline's economics will differ, in either direction.
- No quality verdict. Cost without output assessment cannot tell you which image was good enough to ship.
Summary
The 22x spread between image models is a real number and a mostly useless decision rule. The finding worth acting on is that one parameter inside one model moved costs 35x — more than any model choice on the list — while latency went from 12 seconds to two minutes.
Pick by capability first (reference limits, text output, format), then measure cost per accepted image rather than cost per call, then make sure your draft pass is not running at production quality. The cheapest model on the board and the most expensive one differ by 22x; careless configuration can cost you more than either.
For the models themselves, see GPT Image 2.5, Midjourney, and Google Pics; for the full comparison landscape, the best AI image generators roundup.
Related Articles
Keep reading