An image generation API turns a text prompt into an image over HTTPS. Current per-image prices run from under a cent to $0.21 depending on model and quality tier. Below: exact prices, working code, a capability matrix and the licensing fine print, for every model that matters in 2026.
From ~$0.003 per image via fal, Together or Replicate
Logos, posters, UI copy that actually spells
Multi-turn edits, masks, reference images
Self-host, fine-tune, own the stack
The money table
List prices from provider docs and public aggregator rates, checked July 2026. Where a model has quality tiers, all tiers are shown; the 35x spread between gpt-image-2 low and high is the kind of thing vendors don't put in headlines.
| Model | Price per image | Max resolution | Edit / inpaint | Text rendering | Best at | Access |
|---|---|---|---|---|---|---|
| gpt-image-2 (OpenAI) | $0.006 low · $0.053 med · $0.211 high | 4K-class (3840x2160) | Yes: masks, refs, multi-turn | Strong | Instruction-following, iterative editing | OpenAI API |
| gpt-image-1-mini | ~$0.005 | 1024-class | Yes | Decent | High-volume drafts on OpenAI stack | OpenAI API |
| Imagen 4 (Google) batch = 50% off | $0.02 Fast · $0.04 Standard · ~$0.12 Ultra | 2K | Partial (via Gemini) | Excellent | Text-in-image, speed at quality | Gemini API, Vertex |
| Gemini 3 Pro Image (Nano Banana Pro) | ~$0.039 to 0.24 by tier | 4K | Yes: semantic masks, 14 ref images | Excellent | Complex composition, grounded generation | Gemini API |
| FLUX 2 pro / 1.1 pro (BFL) price-performance king | $0.03 to 0.06 pro · ~$0.003 schnell | 2K (4K upscale) | Yes: Kontext editing | Good | Photorealism per dollar | BFL API, fal, Replicate, Together |
| Ideogram 3 | $0.02 to 0.10 | 2K | Yes | Excellent | Typography, posters, logos | Ideogram API, aggregators |
| Stable Diffusion 3.5 open weights | $0.012 to 0.04 hosted | 2K | Yes | Fair | Self-hosting, fine-tuning, control | Stability API, aggregators, self-host |
| Recraft V3 | $0.03 to 0.09 | 2K + true SVG | Yes | Excellent | Brand design, vector output | Recraft API, aggregators |
| Seedream 4 (ByteDance) | ~$0.02 to 0.04 | 4K | Yes | Good | Value 4K generation | BytePlus, aggregators |
| Aggregators (fal · Replicate · Together) | $0.003 to 0.06 by model | Model-dependent | Model-dependent | Model-dependent | One key, model routing, often cheapest | Their unified APIs |
Integration
Image APIs are synchronous at standard sizes: one request, one image back in 2 to 15 seconds. Here is the task "generate a 1024x1024 product photo" against our unified endpoint. The model is a string; swapping providers is a config change.
curl -X POST \ https://api.aiimagegenerationapi.com/v1/images \ -H "Authorization: Bearer $KEY" \ -d '{ "model": "flux-2-pro", "prompt": "studio photo of a ceramic mug on walnut desk, soft window light", "size": "1024x1024" }' # → { "url": "https://cdn...", # "model": "flux-2-pro", # "cost_usd": 0.03 }
import requests img = requests.post( f"{BASE}/v1/images", headers={"Authorization": f"Bearer {KEY}"}, json={ "model": "gpt-image-2", "prompt": prompt, "size": "1024x1024", "quality": "medium" } ).json() print(img["url"], img["cost_usd"])
const res = await fetch( `${BASE}/v1/images`, { method: "POST", headers: { Authorization: `Bearer ${KEY}` }, body: JSON.stringify({ model: "imagen-4-fast", prompt, size: "1024x1024" }) }); const { url, cost_usd } = await res.json();
Production notes: request 2048-class or 4K only when the use demands it, since price scales with pixels on most providers. Cache aggressively; the same prompt regenerated is money burned. And log per-request cost from day one, because "which feature is spending our image budget" is the first question finance asks.
Beyond generation
Raw text-to-image is table stakes. The differentiators in 2026 are edit operations and reference-image control:
| Capability | What it does | Strongest options |
|---|---|---|
| Inpainting / masks | Regenerate only a masked region: swap a product, fix a hand | gpt-image-2, FLUX Kontext, SD 3.5 |
| Instruction editing | "Make the background white" on an existing image, no mask needed | gpt-image-2, Gemini 3 Pro Image, FLUX Kontext |
| Reference images | Lock a face, product or style across generations | Gemini 3 Pro Image (14 refs), gpt-image-2, Seedream 4 |
| Multi-turn editing | Conversational refinement: each request edits the last result | gpt-image-2, Gemini 3 Pro Image |
| Text-in-image | Legible typography: posters, packaging, UI mockups | Imagen 4, Ideogram 3, Recraft V3 |
| Vector / SVG output | True scalable vectors, not raster tracings | Recraft V3 (unique at production quality) |
| Fine-tuning / LoRA | Train the model on your product or style | FLUX dev, SD 3.5 (open weights), via fal or Replicate |
| 4K generation | Native 3840px-class output without upscaling | Gemini 3 Pro Image, gpt-image-2, Seedream 4 |
Cost at scale
Standard-quality 1024-class images, list prices, no committed-use discounts. This is the table to screenshot for the budget meeting.
| Monthly volume | FLUX schnell (~$0.003) | gpt-image-2 low ($0.006) | Imagen 4 Fast ($0.02) | FLUX 2 pro ($0.03) | gpt-image-2 high ($0.211) |
|---|---|---|---|---|---|
| 1,000 images | $3 | $6 | $20 | $30 | $211 |
| 10,000 images | $30 | $60 | $200 | $300 | $2,110 |
| 100,000 images | $300 | $600 | $2,000 | $3,000 | $21,100 |
The 70x spread between the cheapest and priciest cell is the whole argument for model routing. The pattern that works: generate drafts and thumbnails on a schnell-class model, publish hero assets from a pro-class one, and reserve premium tiers (gpt-image-2 high, Imagen Ultra) for images where text legibility or fine detail is the product. Google's batch API halves Imagen prices for non-realtime workloads; use it for anything a queue can absorb.
The fine print
Hosted closed models (gpt-image, Imagen, Ideogram, Recraft): outputs are yours to use commercially on paid tiers. You own what you generate to the extent the law allows; note that purely AI-generated images may not qualify for copyright protection in the US, which matters if exclusivity is the point.
Open-weight models: read the specific license. Stable Diffusion 3.5 is free under Stability's Community License up to $1M annual revenue, then requires an enterprise deal. FLUX schnell is Apache 2.0 (fully free), but FLUX dev is non-commercial without a paid license key from Black Forest Labs. Shipping FLUX dev output commercially from a self-hosted box without that key is the most common licensing mistake in this space.
Client work: hosted-API outputs can be delivered to clients on all major providers. Keep generation records; ad platforms and stock sites increasingly require AI-provenance disclosure (C2PA metadata), and Meta and Google both auto-label detected synthetic media.
Use-case router
| You're generating | Use | Why |
|---|---|---|
| Ecommerce product shots | FLUX 2 pro, gpt-image-2 | Photorealism plus reference-image support keeps the product accurate. |
| Blog and social heroes at volume | FLUX schnell, Imagen 4 Fast | Cents per image; quality ceiling irrelevant at feed sizes. |
| Posters, packaging, anything with words | Imagen 4, Ideogram 3 | The only models where typography reliably spells. |
| Brand assets and logos | Recraft V3 | True SVG output; hand it straight to the design tool. |
| User avatars / personalization | FLUX dev + LoRA via fal or Replicate | Fine-tune once, generate per-user at open-weight prices. |
| Iterative creative tooling | gpt-image-2, Gemini 3 Pro Image | Multi-turn editing means users refine instead of re-rolling. |
FAQ
Several providers offer free tiers or trial credits: hosted open-weight models are the cheapest real path at roughly $0.003 per image, which is effectively free at prototype volume. Fully unlimited free APIs either watermark output, rate-limit hard or monetize your data. Our unified API includes a free tier at launch; join early access above.
FLUX schnell hosted on fal, Together or Replicate runs around $0.003 per 1024-class image, with gpt-image-1-mini (~$0.005) and gpt-image-2 low ($0.006) close behind. At 100k images per month the spread between cheap and premium tiers is $300 versus $21,000, so tier routing matters more than picking one "cheap" vendor.
Most providers price per image, scaled by resolution and quality tier. OpenAI internally meters image tokens (which is why gpt-image-2 prices vary by quality: $0.006 to $0.211), and Google halves Imagen prices through its batch API. Aggregators normalize everything to a flat per-image rate, which makes cost forecasting simpler.
Yes on paid tiers of every major hosted provider. The traps are open-weight licenses (FLUX dev requires a paid license for commercial use; SD 3.5 is free only under $1M revenue) and copyright: purely AI-generated images may not be copyrightable in the US, so competitors can legally reuse your generated assets.
Imagen 4 and Ideogram 3 lead for legible typography, with Recraft V3 strongest when the text is part of a designed asset like a logo or poster. Classic diffusion models still garble long strings; if words are central to the image, this single capability should drive your model choice.
Yes, and it is the fastest-moving capability. gpt-image-2 and Gemini 3 Pro Image accept plain-language edit instructions on uploaded images, FLUX Kontext and SD 3.5 support mask-based inpainting, and reference-image inputs let you lock a face or product across generations. See the capability matrix above.
Aggregators give you one key across dozens of models, often at or below first-party pricing, with the freedom to reroute when the leaderboard flips. Go direct when you need the newest checkpoints on day one, enterprise SLAs or negotiated volume rates. Most product teams are better served by keeping the model name a config string.
1024x1024 is the standard unit. Gemini 3 Pro Image, gpt-image-2 and Seedream 4 generate native 4K-class output; most others top out at 2K with optional upscaling. Price scales with pixels, so generate at the size you will actually serve.
Yes: FLUX dev/schnell and SD 3.5 run on a single 24GB-plus GPU. You gain fixed costs, privacy and fine-tuning freedom, and take on ops, safety filtering and license compliance (FLUX dev needs a commercial key). Break-even versus hosted APIs typically lands in the tens of thousands of images per month.
Unified schema, per-image billing with cost returned on every response, and model routing across everything in the table above. Early access is open.