Image generation API pricing, compared per image

There is no single way to charge for a generated picture, and that is the main reason image prices are hard to compare by reading vendor pages. The simplest model is a flat fee per image, used by Stability AI, Recraft and xAI models. You send a prompt, you get a picture, and the bill is the same each time.

OpenAI and Google bill image output as tokens instead. A GPT Image model produces a number of image output tokens that depends on the size and the quality setting, and each token is charged at the model's image output rate. The per-image figure you see on OpenAI's pricing page is that token count multiplied by the rate, calculated for a few common sizes. Google works the same way for its Nano Banana models, where a 1K image is a fixed 1,120 tokens and larger sizes use more. A third group, led by Black Forest Labs, charges per megapixel, so a 1024 by 1024 image, which is just over one megapixel, costs the quoted rate and a larger canvas costs proportionally more.

On top of the base rate come options. Quality tiers (low, medium, high) change the price of one model by a factor of several. Size options such as 2K and 4K raise it further. Some vendors charge extra for vector output, for style presets or for editing, where each reference image sent in is billed as input. A few offer an asynchronous batch endpoint at half price. The calculator above lets you set quality, size and batch mode so the ranking reflects the option you actually plan to use.

Every row is reduced to the price of one image at 1024 by 1024 pixels at the default setting: medium quality for OpenAI, the 1K size for Google, one megapixel for Black Forest Labs, and the standard or default tier for everyone else. Token prices are converted using the vendor's own stated token count for that size, and per-megapixel prices are used as quoted because a 1024 by 1024 canvas is treated as one megapixel.

Quality and size are separate price lists. Choosing low or high quality shows each model's tier price at 1K, and choosing 2K or 4K shows its size price at default quality. A model that does not publish the combination you picked is left out of the ranking and counted in the line under the controls, instead of being estimated. Combining a non-default quality with a non-default size returns no price for the same reason: no vendor publishes that grid in a form we can verify. Batch mode swaps in the batch price only at the default quality and 1K size, and only for models that publish one; currently 4 of 35 do.

Monthly cost is the per-image price multiplied by images per day and by 30 days. It excludes retries, failed generations that are still billed, input images for editing, and any plan or platform fee.

Generating 100 images a day for 30 days is 3,000 images a month. At the default setting, the cheapest model in the table, Black Forest Labs FLUX.1 [schnell], charges $0.003 per image, which is $9.00 a month. The median model charges $0.035 per image, or $105 a month. The most expensive, Google Nano Banana Pro, charges $0.134 per image, or $402 a month.

That is a gap of about 45 times between the cheapest and the most expensive option for an image of the same pixel size. Quality is not part of that comparison. A low-cost model and a premium one do not produce interchangeable pictures, and we do not score output quality here, so treat the cheapest row as a floor for budgeting rather than a recommendation.

Open-weight models show a different pattern. The same weights are sold by several hosts, and among the 6 open-weight rows the price runs from $0.003 (Black Forest Labs FLUX.1 [schnell]) to $0.065 (Stability AI SD 3.5 Large). If you can run the weights yourself on rented GPUs the effective cost can be lower again, but that depends on utilization and is outside the table.

Real usage rarely equals the number of images you keep. Designers regenerate several candidates for each image they use, so a retry factor of two to four is common in creative workflows, and the monthly figure scales with it. Set the images-per-day field to your generation count, not your delivered count.

Output quality is not scored. A model that costs a third as much may follow prompts less closely, render text poorly or look less polished, and none of that is visible in a price. Pick candidates from this table by budget, then compare their output on your own prompts before committing.

10 of the 35 rows carry a third-party price: the figure comes from a host or listing rather than the model maker, so it can be out of date or include the host's margin. The Confidence column marks every row, and the Verified only filter removes third-party and reseller prices. For the newest OpenAI models the per-image figures depend on token counts published by third-party hosts, which match the LiteLLM entries we cross-check against.

Model status changes quickly in this category, and we state what we could confirm and what we could not. Google's documentation shows that Imagen is no longer offered in the Gemini API, and that gemini-2.5-flash-image was shut down on October 2, 2026. Third-party sources report that the DALL-E 2 and DALL-E 3 endpoints were retired on May 12, 2026, and that gpt-image-1.5 and gpt-image-1-mini are scheduled to shut down on December 1, 2026; we have not confirmed either report against OpenAI's own documentation, so check before you build on those models. 3 rows are flagged with a retirement date in the table.

Free tiers, committed-use discounts, enterprise agreements and regional pricing are not included. Bulk upload credits and prepaid bundles are not modeled either.

Prices come from two kinds of source. Where a vendor publishes a price page we read it and record the URL, and those rows are tagged Official. Where we rely on the LiteLLM community price list, the row is tagged LiteLLM. Rows whose price comes from a host, aggregator or other listing are tagged Third-party, and prices charged by a reseller rather than the maker are tagged Reseller. Each row stores its published price basis, source link and the date we last checked it.

The dataset was last updated on 2026-10-24 and refreshes daily. LiteLLM values are re-derived on each run, and a value that falls outside plausible bounds for an image price is dropped instead of published. If a price here is wrong, use the contact link at the bottom of the page and include the model and a source.

Not listed: DALL-E 2 and DALL-E 3 (reported retired on 2026-05-12), Imagen (Gemini API) (removed from the Gemini API; Vertex AI price not verified), Gemini 2.5 Flash Image (shut down on 2026-10-02), Midjourney (no official API), FLUX 3 (no price available), GPT Image 1 (LiteLLM lists a retirement date of 2026-10-23), Titan Image Generator v2 (LiteLLM lists a retirement date of 2026-06-30), Nova Canvas (LiteLLM lists a retirement date of 2026-09-30) and MAI-Image-2.5, 2.5 Flash and 2.5 Pro (LiteLLM lists a retirement date of 2026-10-01; the 2.6 models are priced per token only). If you also generate motion, the video generation page prices clips per second on the same plan.