The comparison problem starts before quality
If you are picking an image model, the price column on a model page will not help you. OpenRouter’s image generation comparison makes the reason concrete: the 52 models it routes are not sold in the same unit. FLUX.2 bills per megapixel. Gemini, OpenAI, and Microsoft MAI bill per token. Grok, Recraft, Riverflow, Qwen, and Seedream bill per image, at a rate that shifts with quality and resolution.
Each model page is correct in its own unit. That is exactly why you cannot read two of them side by side and decide anything.
What the measured run actually showed
OpenRouter sent the same prompt to 20 models at the same size on 11 September 2026, then read the usage.cost field from each response. One image at default settings billed between $0.006 and $0.134 — a 22x spread. The billing units were mixed all the way up that range, so the unit itself tells you nothing about where a model lands.
Three findings matter more than the table:
- Quality settings dominate on OpenAI models. The same prompt at 1024x1024 went from $0.006 with quality unset to $0.211 at
high. The high call also took 123 seconds against 12 seconds for the default. - Listed rates and billed cost diverge. Riverflow 2.5 Pro listed $0.13 per image at default resolution and billed $0.064. FLUX.2 Flex listed $0.06 per megapixel and billed $0.05. Krea 2 Medium published no pricing record and billed $0.03.
- Defaults can surprise you. Both Seedream models returned 2048x2048 when resolution was left unset, and Seedream 5.0 Pro billed its high-resolution rate rather than its base rate.
OpenRouter’s own guidance is to read usage.cost rather than multiply a listed rate by an estimate. It also notes that prices and models change often, so any figure more than a month old is a hint rather than a budget line. You can re-pull current state with GET /api/v1/images/models and GET /api/v1/images/models/{model}/endpoints.
Capability ceilings rule models out before price does
Cost is one axis. Two others eliminate candidates outright.
The first is reference images. Grok Imagine takes three; Gemini 3.x takes fourteen. A six-image brand kit rules Grok out before you compare anything else. The second is output modality: most models return only image bytes, while Gemini image models and OpenAI’s gpt-5-image models can return written text in the same reply.
Format handling is a smaller trap with the same shape. FLUX.2, Grok, and Seedream returned JPEG; OpenAI, Qwen, Microsoft, and Krea returned PNG; Sourceful and Recraft returned WebP, and Recraft’s vector model returned SVG. The three Gemini models tested do not accept output_format and did not all return the same format, so do not hard-code a file extension. Read the media_type field instead.
On reproducibility, seed is absent from every Gemini, Grok, OpenAI, Riverflow, Recraft, and MAI image model in the catalog. Where it exists on FLUX.2 Klein, the same seed reproduced the image byte for byte. OpenRouter only tested Klein, so treat other seed-capable models as untested.
Picking by job instead of by leaderboard
OpenRouter’s shortlist is job-shaped: start with openai/gpt-image-2 at $0.006 for short prompts, reach for sourceful/riverflow-v2.5-pro at $0.064 when the image is hard to get right, google/gemini-3-pro-image at $0.134 when the reply needs words as well as a picture, and recraft/recraft-v4.1-vector at $0.08 when you need editable SVG.
Design Arena ratings are shown on each model page, and Riverflow 2.5 Pro held the highest rating on both the image and image-editing boards when pulled on 11 September 2026. But the top-rated model billed $0.064 in this run and the lowest-rated one billed $0.014. Elo reflects what voters preferred on their prompts, not yours. Shortlist from the board, then decide from your own prompts.
The practical takeaway: generate one image per candidate, read usage.cost off the response, and budget from an average of several calls. Two identical Riverflow 2.5 Fast requests billed $0.017623 and $0.017639. If you are consolidating image, speech, and video calls behind one request shape, the same discipline applies — see our earlier post on OpenRouter’s multimodal API consolidation.
One limitation worth stating plainly: this is a single prompt at a single size on one day. It tells you how to measure, not what your bill will be.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.