Ask "what does an AI image cost?" and you get three incompatible answers, because the three dominant platforms sell three different things: Midjourney sells subscription time, DALL-E sells generations through API and platform credits, and Stable Diffusion sells nothing at all — you bring the hardware. This comparison lays out the three cost models, the arithmetic that converts each into a cost per usable image, and a worked example you can re-run with current prices.
The Three Cost Models
Midjourney — subscription allowance. Plans are flat monthly fees that include a ration of "fast GPU hours" (with slower "relaxed" generation treated differently by tier). Your cost driver is not images but GPU time consumed: high resolutions, upscales, and variations burn the allowance faster. Effective cost per usable image = subscription ÷ usable images actually produced that month.
DALL-E (OpenAI) — per-generation pricing. Accessed via API or bundled platform credits, cost attaches to each generation, with price steps by model, resolution, and quality setting. Effective cost = price per generation × generations attempted — there is no allowance, so spend scales linearly from the first image.
Stable Diffusion — self-hosted compute. The model weights carry no per-image provider fee (licensing terms vary by model version — check the specific release). Costs relocate to GPU hardware or cloud-GPU rental, electricity, and your setup and maintenance time. Marginal cost per image trends toward the electricity-and-minutes floor, but only after real fixed investment.
The Number That Decides Everything: Attempts per Usable Image
Nobody ships their first generation. Briefs get explored, hands come out wrong, compositions get rejected — a realistic workflow needs 3–5 generations per usable image. Every model comparison that ignores the discard rate understates cost by exactly that factor:
Worked Example (Illustrative Figures — Reprice Before Use)
Target: 500 usable images in a month, at 3 attempts per usable = 1,500 generations. The prices below are round illustrative numbers chosen to expose the arithmetic, not current quotes — substitute live prices from each provider before budgeting.
| Model | Illustrative pricing | Monthly cost for the workload | Cost per usable image | | --- | --- | --- | --- | | Subscription (Midjourney-style) | 30 flat, if the GPU-hour allowance covers 1,500 generations | 0.04 per generation | 1,500 × 60 | 1,200 GPU over 24 months (8 electricity | ≈ 0.116 |
The table teaches the real lesson: at this scale the three models land within cents of each other, and the ranking flips with volume and discard rate. Drop to 50 usable images a month and the subscription's flat fee becomes 0.12; rise to 5,000 usable images and the API bill grows linearly to 0.012 per image — if one GPU's throughput keeps up. Upscaling, inpainting passes, and editing steps count as further generations in every model; price the whole workflow, not the first render.
Choosing by Situation
Choose the subscription for predictable, heavy creative exploration by a person: flat cost, fast iteration, no infrastructure — and read the commercial terms, since higher-revenue companies are required onto higher tiers by Midjourney's license. Choose per-image API pricing for automation and variable volume: product pipelines, programmatic generation, and anything that must scale from zero without a monthly commitment. Choose self-hosting for sustained high volume, privacy-sensitive work, or fine-tuned custom models, where you control the stack and the marginal cost floor — accepting that "free" software still bills you in hardware, power, and engineering hours.
Whichever route, verify current pricing on the provider's official page at decision time: AI image pricing has been repriced repeatedly and any undated figure (including the illustrative ones above, by design) ages in months. For text-model spend, the companion LLM API Cost Calculator applies the same discipline to tokens.
Frequently Asked Questions
Is Stable Diffusion really free?
The weights for many releases can be downloaded without a per-image charge, but generation is not free: it consumes GPU hardware you bought or rented, electricity, and your time, and specific model versions carry license conditions for commercial use. It is "no vendor bill," not "no cost."
Why do comparisons online give one flat price per image?
Because they divide by generations, not usable images, and they freeze one day's prices. Apply the attempts-per-usable multiplier and refresh prices at budgeting time and the same comparison becomes honest — usually 3–5× the headline figure per finished asset.
Which is cheapest for a small business?
At low volume, per-image API pricing almost always wins: no subscription floor, no hardware. The crossover arrives with sustained monthly volume or a high discard rate, at which point a subscription's flat allowance or self-hosting's diluted fixed cost pulls ahead — run the worked example's arithmetic with your real volume.