Back to News

What OpenRouter’s 20-Model Image Test Means for Visual Creators

OpenRouter compared 20 image generation models with the same prompts and found a 22x spread in measured per-image cost. For creators, the takeaway is practical: compare models by actual usage.cost, not list pricing, and choose by task constraints such as readable text, reference-image limits, SVG output, or text-plus-image replies.

AiPixSeptember 19, 2026

OpenRouter’s comparison of 20 image generation models is useful because it avoids the usual pricing mismatch problem. According to OpenRouter, different models charge by different units: FLUX.2 by megapixel, Gemini, OpenAI and Microsoft MAI by token, and Grok, Recraft, Riverflow, Qwen and Seedream by image, with rates affected by quality and resolution. OpenRouter therefore sent the same prompt to 20 models and read the actual usage.cost returned by each response.

Source Facts

OpenRouter says every price in its article came from a real generation measured on 2026-09-11. The default 1:1 prompt was: “A studio packshot of a glass bottle, hard key light, no text.” Across the 20 tested models, measured cost ranged from $0.006 to $0.134 per image, a 22x difference.

The cheapest measured model was openai/gpt-image-2, at $0.006 for the short prompt and $0.0135 for a longer coffee-bag prompt. OpenRouter also found that sourceful/riverflow-v2.5-pro cost $0.064 in the default image table and led the Design Arena image and image-editing leaderboards when OpenRouter pulled benchmark scores on 2026-09-11. google/gemini-3-pro-image cost $0.1344 in the default table, while recraft/recraft-v4.1-vector cost $0.0800 and returned SVG.

OpenRouter’s quality-setting result is especially important for budgeting. On openai/gpt-image-2, keeping quality unset or setting it to low produced 196 image tokens and a $0.00599 charge. Setting quality to high produced 7,024 image tokens and a $0.21083 charge. OpenRouter also reports that the high-quality call took 123 seconds, compared with 12 seconds for the default call.

For text rendering, OpenRouter tested six models on a matte black coffee bag with two lines of label text. Five of the six rendered both lines correctly. The exception was black-forest-labs/flux.2-klein-4b, which rendered the brand name correctly but misspelled the second line as “SINGLE ORISION.”

Reference-image limits also varied. On 2026-09-11, OpenAI image models accepted 16 reference images, Gemini 3.x image models and Seedream 5.0 accepted 14, sourceful/riverflow-v2.5-pro accepted 10, FLUX.2 Flex, Pro and Max accepted 8, FLUX.2 Klein and Qwen Image 3 accepted 4, Grok Imagine accepted 3, and Recraft, Krea and MAI Image accepted 1.

AiPix Analysis

For visual creators, the practical lesson is not “pick the cheapest model.” It is “test the exact workflow.” A product-shot workflow with no readable text may tolerate a cheaper or more stylized model. Packaging, UI mockups, posters and ads need stricter text checks. Brand-consistent editing needs enough reference-image capacity, plus visual review against the source image.

OpenRouter’s findings also reinforce why a canvas-based workflow matters. In AiPix terms, generation cost is only one part of the creative decision. Creators need to compare outputs, preserve reference intent, remove or enhance backgrounds when needed, and move promising results into one-click paths such as portraits, ID photos or wedding concepts. The model choice should follow the asset’s job: bulk variants, editable SVG labels, reference-based edits, or replies that need both image and text.

One operational rule stands out: do not budget from pricing tables alone. OpenRouter recommends reading usage.cost from real calls, because listed rates and actual measured charges can differ. For teams producing images at scale, a small prompt change, a quality setting, or a reference edit can meaningfully change the bill.

Sources

  1. OpenRouter 实测 20 个图像生成模型的成本、编辑与质量OpenRouter:Announcements(RSS)

Continue in AiPix

Turn this update into a practical image or canvas workflow.

Open AI Canvas