Ten thousand images a month is a working budget for a mid-sized e-commerce catalog, and in 2026 the same workload prices out between $90 and $2,110 — a 23x s...
Ten thousand images a month is a working budget for a mid-sized e-commerce catalog, and in 2026 the same workload prices out between $90 and $2,110 — a 23x spread — depending on which API you call. Nothing in the output justifies most of that gap. The cheap end is an open-weight model running in someone else's data center; the expensive end is a token meter charging $0.211 per image at one particular quality setting.
AI image generation pricing in 2026 runs on two systems that reward different habits. OpenAI and Google's Nano Banana line bill by token, which turns every quality knob into a direct budget line. Black Forest Labs and Google's Imagen bill per image at fixed resolution tiers, which makes the math boring and predictable. Stability still sells credits — and still hands new accounts only 25 of them, roughly three images on its flagship, as a trial.
This guide covers the four providers developers actually bill against: OpenAI's gpt-image family, Google's Imagen 4 and Nano Banana stack, Black Forest Labs' FLUX 3 lineup, and Stability's Stable Diffusion 3.5 API. Every number comes from official pricing pages read on November 20, 2026. Where a figure is a calculator estimate rather than a published rate, the text says so. If you want the consumer side — Midjourney subscriptions, Canva credits, editor seats — that is a different market, covered in our image generation tools comparison.
AI image generation pricing at a glance
| Provider | Entry 1K image | Standard 1K image | Premium 1K image | 4K-class image | Billing model |
|---|---|---|---|---|---|
| OpenAI | $0.006 (2.5 low, calc.) | $0.013 (2.5 medium, calc.) | $0.211 (gpt-image-2 high) | Not offered (1,536 px max) | Tokens, $30/M output |
| Google — Imagen 4 | $0.02 (Fast) | $0.04 | $0.06 (Ultra) | $0.04 + $0.06 upscale | Per image |
| Google — Nano Banana | $0.034 (2.1) | $0.067 (2) | $0.134 (Pro, 1K/2K) | $0.113 (2.1) to $0.24 (Pro) | Tokens, $30–$120/M output |
| Black Forest Labs | $0.009 (FLUX 3 Klein) | $0.048 (FLUX 3) | $0.10 (FLUX 3, 2K) | $0.036 (Klein) to $0.607 (FLUX 3) | Per image, by resolution tier |
| Stability | $0.025 (SD3.5 Flash) | $0.04 (Large Turbo) | $0.08 (Stable Image Ultra) | Separate upscale products | Credits, 1 credit = $0.01 |
Three structural facts sit behind that table. The cheapest current-generation 1K image on the market costs $0.009 — FLUX 3 Image Klein — and it is not a toy: Black Forest Labs ships the same weights for self-hosting. OpenAI's per-image figures are estimates, not rates, because the actual AI image generation cost per image moves with quality, size, and even aspect ratio. And 4K is where pricing turns strange: BFL charges $0.607 for its own 4K tier — 17 times what its Klein variant asks — while Google's Nano Banana 2.1 produces true 4K for $0.113.
How each provider meters an image
OpenAI: the gpt-image-2 pricing ladder, token by token
OpenAI's Standard rates are identical across gpt-image-2, 2.5, and 2.6: $5 per million text-input tokens, $8 per million image-input tokens, $30 per million image-output tokens, with cached image input at $2. What varies is how many tokens a single image burns. For gpt-image-2, OpenAI's published cost table reads $0.006 / $0.053 / $0.211 for low, medium, and high at 1024×1024, and $0.005 / $0.041 / $0.165 at 1536×1024. Two details are easy to miss: landscape orientations bill roughly 20% less than squares, and the input side is not free — an edit request bills your reference image as image input at $8 per million tokens.
The 2.5 generation (Flare and Sunburst) and the current 2.6 models share those token rates. Independent readings of OpenAI's cost calculator from October 2026 agree on the token counts for 2.5 at 1024×1024: 196 tokens at low (about $0.006), 439 at medium ($0.013), 1,756 at high ($0.053), 3,122 at xhigh ($0.094), and 7,024 at max ($0.211). OpenAI itself warns that "the same quality label does not imply the same image quality or response time across models," and its model pages caution that the calculator does not estimate 2.5 token consumption — treat these as estimates and read usage.output_tokens from your own responses. The gpt-image-1-mini pricing lane stays cheapest at the token level: $2 text in, $2.50 image in, $8 out, though OpenAI publishes no per-image table for it.
Google: Imagen 4 pricing by the image, Nano Banana by the token
Imagen 4 pricing is the simplest rate card in the industry: $0.04 per image, $0.02 for Imagen 4 Fast, $0.06 for Imagen 4 Ultra. Upscaling an Imagen 4 output to 2K, 3K, or 4K costs a flat $0.06 on top of the generation. That is the entire published price list.
The Nano Banana stack bills tokens instead. Nano Banana Pro (Gemini 3 Pro Image) charges $120 per million image-output tokens, which works out to $0.134 for 1K or 2K output (1,120 tokens) and $0.24 for 4K (2,000 tokens). Nano Banana 2 (Gemini 3.1 Flash Image) runs $60 per million: $0.067 at 1K, $0.101 at 2K, $0.15 at 4K. Nano Banana 2.1 halves that again: $30 per million image-output tokens, which means $0.034 at 1K, $0.05 at 2K, and $0.113 at 4K. Input images cost tokens too — 560 per reference image on Pro, 1,120 on 2.1 — so multi-reference edits are cheaper than on OpenAI but not free. On community leaderboard votes from September 2026, Nano Banana 2 trails OpenAI's Sunburst (1,261 vs 1,421), which is the honest context for its price advantage. The models live inside Gemini if you already pay Google for tokens.
Black Forest Labs: FLUX 3 API pricing by resolution tier
FLUX 3 API pricing is flat per image, and the tier you request — not the aspect ratio — sets the bill: $0.041 at 768×768, $0.048 at 1K, $0.07 at 1.5K, $0.10 at 2K, and $0.607 at 4K. Reference images are included at no charge, failed generations are not billed, and every aspect ratio inside a tier costs the same. FLUX 3 Image Klein, a 5.6B-parameter model Black Forest Labs says runs on a single consumer GPU, cuts every tier by roughly 80%: $0.008, $0.009, $0.011, $0.014, and $0.036. The older FLUX.2 family (pro, flex, max, klein) remains fully supported for production work, and third-party hosts fill the gaps — Replicate lists FLUX 1.1 Pro at $0.04 per image, a rate our GPU cloud pricing review verified in November.
Stability: stable diffusion API pricing in credits
Stable diffusion API pricing runs on credits: one credit equals $0.01, billed per generation. SD3.5 Flash costs 2.5 credits ($0.025), SD3.5 Medium 3.5 ($0.035), Large Turbo 4 ($0.04), SD3.5 Large 6.5 ($0.065), Stable Image Core 3 ($0.03), and Stable Image Ultra 8 ($0.08). Legacy SDXL starts at $0.009 — still the cheapest published rate on any mainstream API, for 2023-era quality. The free tier is thin and one-time: 25 credits for new accounts, about three images on the Ultra flagship, and credits issued after September 30, 2026 expire a year later under Stability's terms. Edit and upscale operations bill separately — inpainting 5 credits, Fast Upscaler 2, Creative Upscaler 60. The open weights remain the real offer: the Community License permits commercial self-hosting under $1 million in annual revenue, and the Stable Diffusion ecosystem is still the largest self-host install base in the field.
The relabeled ladder: why gpt-image-2.5 pricing looks 4x cheaper
Here is the change none of the release notes explained plainly. gpt-image-2.5 uses the same token rates as gpt-image-2 but spends different token budgets under the same quality labels. At 1024×1024, "high" on 2.5 consumes 1,756 output tokens — the budget gpt-image-2 spent at "medium." A migration that keeps quality set to "high" therefore drops from $0.211 to about $0.053 per image, a 4x cut that is really a relabel; the old high budget now lives at "max" ($0.211). Two 2.5 rungs — medium and xhigh — have no equivalent on gpt-image-2 at all.
The practical consequence: never compare quality labels across model generations. Compare token counts, then read actual usage from responses. Sunburst and Flare bill identically, and the difference is workflow, not wallet: OpenAI's docs put Sunburst on "editing precision" and Flare on "fast, high-quality everyday generation," while one operator reporting roughly 50,000 images on Hacker News measured average latency dropping from about 104 seconds on gpt-image-2 to 35–40 seconds on 2.5. gpt-image-2.6 adds a complexity setting with no separate fee. Arena crowd votes from September 2026 rank Sunburst first (1,421) and Flare second (1,399) on text-to-image — third-party data, but the largest public sample available.
Worked bills: what AI image generation pricing means per month
Rates hide budgets, so here are four real workloads priced at official rates, with text-input overhead excluded for clarity (a 150-token prompt adds about $0.75 per thousand images).
| Workload | Lowest-cost route | Monthly bill | Premium route | Monthly bill |
|---|---|---|---|---|
| 10,000 catalog shots, 1K | FLUX 3 Klein @ $0.009 | $90 | Nano Banana Pro @ $0.134 | $1,340 |
| 10,000 hero images, 1K | gpt-image-2.5 high @ about $0.0527 | $527 | gpt-image-2 high @ $0.211 | $2,110 |
| 10,000 renders at 4K | FLUX 3 Klein @ $0.036 | $360 | FLUX 3 @ $0.607 | $6,070 |
| 50,000 thumbnails, 1K | gpt-image-2.5 medium via Batch @ about $0.0066 | $329 | SD3.5 Flash @ $0.025 | $1,250 |
Two rows deserve a second look. The hero-image row is the same render budget moved from gpt-image-2 "high" to 2.5 "max" — $2,110 either way — while an operator who accepts 2.5's new "high" label pays $527 for images rendered at the old medium budget. The 4K row is the trap from the glance table: BFL's own lineup spans $360 to $6,070 for identical resolution, a 16.9x gap inside one vendor.
Batch modes, streaming, and the fine print
OpenAI's Batch API halves every token rate — $15 per million image-output tokens — on the same 24-hour completion window our LLM API pricing comparison puts to work. That is the cheapest effective rate on the market for patient workloads: 2.5 "medium" through Batch lands near $0.0066 per image, undercutting Klein's published $0.009 if the delay fits your pipeline. Google's Flex mode prices Nano Banana Pro image output at $60 per million tokens — $0.067 for a 1K image, half list — though Nano Banana 2.1 does not support Flex at all. Stability publishes no batch discount, and BFL's per-image tiers are already flat.
Three fine-print items move real money. Streaming previews add 100 image-output tokens per partial frame — three partials add $0.009, which is more than the underlying low-quality render itself, so "watch the image draw" UI demos are a luxury. OpenAI applies a 10% regional processing surcharge to models released on or after March 5, 2026. And Imagen's upscale path means a "4K image" can be generation plus a $0.06 operation — $0.10 all-in, the cheapest verified route to a 4K-class image on any API.
Self-hosting: the honest answer, with strings attached
Self-hosting is the fair response to "why pay per image at all," and the open-weight lineup is genuinely good now. FLUX.2 klein ships under Apache 2.0, FLUX 3 Image Klein's weights are licensable through BFL, and SD3.5 Medium runs commercially free below $1 million in annual revenue. Black Forest Labs sells self-host licenses on tiers — Builder for 10K images a month, Platform for 100K — with pricing on application.
Run the breakeven from the API side: FLUX 3 Klein pricing works out to $900 a month for 100,000 images at 1K. No single-GPU rig plus the engineer who babysits it beats $900 on cost alone at that volume — the GPU bill is fixed and utilization is not. Self-hosting wins when volume is steady and high, when outputs must stay on-premise, or when fine-tuning is the product; our GPU cloud pricing comparison covers the hardware math. The middle path is per-image hosting of open weights on aggregators such as fal.ai, which skips the infrastructure bill while keeping per-image economics.
Frequently asked questions
What is the cheapest AI image generation API in 2026?
For published rates, FLUX 3 Image Klein: $0.009 per 1K image, $0.008 at 768px. Runners-up: SDXL at $0.009 on Stability (legacy quality), Imagen 4 Fast at $0.02, and SD3.5 Flash at $0.025. If you can wait 24 hours, OpenAI's Batch lane on gpt-image-2.5 "medium" lands near $0.0066 — a calculator estimate rather than a published rate, so budget with margin.
Does gpt-image-2.5 pricing differ from gpt-image-2?
Token rates: identical. Token budgets per quality label: different. The 2.5 "high" setting spends what gpt-image-2's "medium" spent (1,756 output tokens at 1024×1024), so the same label costs about 4x less — and "max" restores the old high budget at $0.211. Same money, finer ladder.
What does Nano Banana Pro pricing look like per image?
$0.134 per image at 1K or 2K, $0.24 at 4K, from $120 per million image-output tokens. Reference images add 560 input tokens each. Flex mode halves the rate to $0.067 for patient workloads. Pay that premium for grounded, text-accurate renders — not for catalog volume, where it costs 15x Klein for images the community leaderboard scores lower.
Is FLUX 3's 4K tier worth $0.607 per image?
Rarely. The same vendor's Klein tier renders 4K for $0.036 — 16.9x cheaper — and Google's routes undercut both: $0.10 via Imagen 4 plus upscale, $0.113 via Nano Banana 2.1. FLUX 3's 4K earns its price only when a multi-reference edit has to survive at 16 megapixels.
Which API wins for batch image generation?
Catalogs with 24-hour tolerance: OpenAI Batch on gpt-image-2.5 "medium" — about $329 for 50,000 images. Immediate delivery: FLUX 3 Klein at $450 for the same 50,000. Google Flex halves Nano Banana Pro's rate for patient premium work; Stability has no batch lane. On OpenAI, favor 1536×1024 over square — the same quality setting bills roughly 20% fewer tokens.
The bottom line
AI image generation pricing in 2026 rewards two habits: measure your own token counts before trusting any per-image estimate, and match the billing model to the workload. Volume catalogs belong on per-image tiers — Klein at $0.009 or Imagen 4 Fast at $0.02. Quality-critical single shots belong on token meters — 2.5 "high" at about $0.053 or Nano Banana Pro at $0.134 — where you pay only for the renders that matter. Patient pipelines take OpenAI's Batch and Google's Flex at half rate. And the 4K premium is a mirage outside two specific lanes: $0.10 for Imagen 4 plus upscale, $0.113 for Nano Banana 2.1. Prices here were read from official pages on November 20, 2026; image models reprice quarterly, and our AI tool pricing index tracks the drift.