Image generation APIs bill by image, token, megapixel or credit, and the sticker price rarely matches the final invoice. This article compares OpenAI, Google, FLUX and Stability AI prices, shows which options are free, and walks through a hands-on PicassoIA Image tutorial with a cost table for 10,000 images.
Most pricing pages for an image generation API quote one friendly number and bury the rest. A headline rate of a few cents per image grows fast once you ask for higher resolution, rerun a bad attempt, or send reference photos that are billed as input. This article lays out what each major provider charges, which options cost nothing, and how to keep a monthly bill predictable. Prices were checked on official pages and trusted trackers in October 2026, so confirm the live rate before you commit a budget.
Two providers can both advertise "pay as you go" and still bill in entirely different units. Sorting that out first makes every later comparison honest.
There are three billing styles in the wild, and each one changes how you forecast spend:
Google sits in between: its Gemini image models are token based, but the pricing page converts tokens into a per-image figure for each resolution tier.
The sticker price is only the starting point. Four things push the real number up:
💡 Tip: Track cost per accepted image, not cost per request. A model that is twice the price but needs half the retries can be the cheaper choice.
Behind every one of these rates sits GPU time. Providers rent racks of accelerators, and your price is mostly their hourly cost divided by how many images fit into that hour. That is why fast, small models are cheap and large, slow, high-resolution models are not.
Here is the field at a glance. Figures come from each provider's pricing page where it was readable, and from published trackers where it was not. Rows marked "reported" deserve a second look on the vendor's own dashboard before you sign off on a budget.
| Provider and model | Billing unit | Price signal | Free access |
|---|---|---|---|
| OpenAI GPT Image 2 | Tokens | $30 per 1M image output tokens | None listed |
| Google Nano Banana 2 Lite | Per image | $0.0336 at 1K | Testing in AI Studio only |
| Google Nano Banana 2 | Per image | $0.067 at 1K, $0.151 at 4K | Testing in AI Studio only |
| Google Nano Banana Pro | Per image | $0.134 at 1K or 2K, $0.24 at 4K | Testing in AI Studio only |
| Google Imagen 4 | Per image | $0.02 Fast, $0.04 Standard, $0.06 Ultra (reported) | None listed |
| FLUX.2 Klein 4B | Per megapixel | From $0.014 | Open weights |
| FLUX.2 Pro | Per megapixel | From $0.03 | None listed |
| Stability AI Core | Credits | 3 credits, about $0.03 | 25 starter credits (reported) |
| Stability AI Ultra | Credits | 8 credits, about $0.08 | 25 starter credits (reported) |
| PicassoIA Image | API predictions | Listed as free, see below | Yes, for now |
OpenAI publishes token rates rather than a flat per-image price. GPT Image 2, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst all list $8 per million image input tokens, $30 per million image output tokens and $5 per million text input tokens. Cached image input drops to $2.
The budget option is GPT Image 1 Mini at $8 per million image output tokens, roughly a quarter of the flagship output rate. Batch processing is priced at about half of standard rates, which suits overnight jobs that nobody is waiting on.
The catch with token billing is forecasting. A larger canvas or a higher quality setting produces more output tokens, so run a test batch and read the usage numbers before you multiply by your expected volume.
Google gives the clearest per-image table of the big vendors:
The paid image models have no free tier in the API. Free access is limited to testing inside Google AI Studio. You can also try the lighter Nano Banana 2 Lite on PicassoIA without wiring up billing.
Imagen 4 is reported at $0.02 for Fast, $0.04 for Standard and $0.06 for Ultra. Those figures come from third-party trackers, so verify them in your Google Cloud console.
Black Forest Labs states "no subscriptions, no seat fees," and you pay only for what you generate. The FLUX.2 family is metered by megapixel, so every figure below is a starting point:
Batching does not lower these rates, so the savings come from choosing the right tier and the right resolution, not from queueing jobs.
Stability AI sells credits at $0.01 each, and every endpoint deducts a fixed number. Stable Image Core uses 3 credits (about $0.03) and Stable Image Ultra uses 8 (about $0.08). A $10 top-up buys 1,000 credits, and new accounts are reported to receive 25 free credits for testing. If you prefer the open model line, Stable Diffusion 3.5 Large is available on PicassoIA.
"Free" means very different things across vendors. Some give a trial balance, some give a sandbox, and one gives the API itself.
PicassoIA exposes a Replicate-style REST API at https://api.picassoia.com/v1. Its documentation states: "API predictions are currently free. They use no credits." Four models are available through it: PicassoIA Image for text-to-image, PicassoIA Image Editor Pro for editing and combining images, PicassoIA Video and Seedance 2.5 Lite for video.
There are limits to plan around. An account can have up to 5 predictions queued or running at once, shared across all its API tokens and MCP connections. Requests are capped at 10 MB, prompts at 4,000 characters, and jobs time out after three hours.
💡 Check your plan first. The API page says access requires an Infinite plan account, while the pricing page lists API Access and MCP Connections under Pro+, Elite and Infinite. Confirm which tier your account needs before you build on it, and treat "free" as the current policy rather than a permanent promise.
Smaller free offers are worth using for prototypes:
Open weights remove the per-image fee, but not the cost. You pay for a GPU instead. FLUX.1 schnell and FLUX.2 klein are the best-known open-weight options, and FLUX Schnell is a fast, forgiving place to start. Check the license of each variant before you ship it in a commercial product, because terms differ between model sizes.
Self-hosting wins when your volume is steady and high. It loses when your traffic is spiky, because an idle GPU still bills by the hour.
PicassoIA Image is the platform's native text-to-image model. It returns an image in a few seconds, and it is described as unlimited, with no per-image cap and no daily quota.
💡 Tip: Generate two variations per call while you iterate. You compare results side by side without paying a retry penalty.
The same model is reachable by HTTP. Create a prediction, then poll its URL until the status reads succeeded:
curl -s -X POST https://api.picassoia.com/v1/models/picassoia/picassoia-image/predictions \
-H "Authorization: Bearer $PICASSOIA_API_TOKEN" \
-H "Content-Type: application/json" \
-d '{"input": {"prompt": "a lighthouse at sunset, photorealistic", "aspect_ratio": "16:9"}}'
The response returns an id, a status of starting and a urls.get address. Poll that address with the same Authorization header until status becomes succeeded, then read the output field. Because jobs are asynchronous, build your client around create, poll and fetch rather than waiting on one long request.
At 10,000 images a month, the gap between models becomes real money. Using the list prices above at 1 MP or 1K resolution and no retries:
| Model | Price per image | Cost for 10,000 images |
|---|---|---|
| FLUX.2 Klein 4B | from $0.014 | about $140 |
| Imagen 4 Fast (reported) | $0.02 | about $200 |
| FLUX.2 Pro | from $0.03 | about $300 |
| Stability AI Core | $0.03 | about $300 |
| Nano Banana 2 Lite | $0.0336 | about $336 |
| Nano Banana 2 at 1K | $0.067 | about $670 |
| Stability AI Ultra | $0.08 | about $800 |
| Nano Banana Pro at 1K | $0.134 | about $1,340 |
Add a 33% retry penalty and the Nano Banana Pro line climbs past $1,780. Three habits keep that number honest.
Generate the idea at 1 MP, pick the winners, and only then pay for size. A dedicated upscaler turns a good draft into a print-ready file for a fraction of the cost of rendering everything at 4K. Real-ESRGAN and Crystal Upscaler both handle that job on PicassoIA.
Weak prompts cause retries, and retries are the most expensive line on your invoice. A language model can expand a one-line brief into a precise photographic prompt before any image is billed. Claude Sonnet 5 and Gemini 3.5 Flash are both available on PicassoIA for that step. Spending a fraction of a cent on text to save a 3 to 8 cent re-render is a clear win.
Cache and reuse seeds. Store every accepted image with its prompt and seed. If a customer asks for the same product shot twice, serve the stored file instead of paying again. When a seed is supported, a locked seed also lets you tweak one detail without losing the composition.
The cheapest model on paper is not always the right one. Match the tool to the job:
A small shop might need 500 product images a month. At FLUX.2 Pro rates that is roughly $15, which is trivial next to a photographer's day rate. Here consistency matters more than cost, so lock your prompt template and seed, then spot check each batch.
Interactive products care about latency and concurrency as much as price. The PicassoIA API allows 5 simultaneous predictions per account, so a busy app needs a queue and a retry policy. Paid providers raise limits as your spend grows, which is a real advantage once traffic climbs.
Numbers on a pricing page only tell you so much. The fastest way to judge an image generation API is to run your own prompts and count how many results you would actually publish. Open PicassoIA Image, generate a handful of variations of a scene from your own project, and compare them with the paid models above. Then try PicassoIA Image Editor Pro on one of your own photos, and browse the full catalog at picassoia.com/en/all-models. Experiment freely: the cheapest way to find the right model is to test it on the work you actually need done.
