AI Image Model Pricing: Every Official Rate, Normalized Per Image (August 2026)
Last updated August 7, 2026
Ten image-model families bill in four incompatible units: per image (Imagen, Qwen, Ideogram, Recraft, Luma), per megapixel (Flux 2), per token (GPT Image, MAI), and per edit (Reve, $0.01). Normalized to one 1024×1024 image, the market clears at $0.03–$0.04 — Flux pro, Qwen, Ideogram Turbo, Recraft V4.1, Imagen Standard, and Luma uni-1.1 all land there. Imagen 4 Fast's $0.02 dies with the Aug 17, 2026 shutdown. All ten families run inside the invideo agent.
Updated August 2026
AI image model pricing looks like a solved comparison — every vendor publishes numbers — yet it is one of the most misleading spec sheets in AI, because the vendors never agreed on a billing unit. OpenAI and Microsoft meter images in tokens. Black Forest Labs bills by the megapixel. Google, Alibaba, Ideogram, Recraft, and Luma bill per image — at different resolutions. Reve bills editing by the operation, at one cent. "A $0.03 image" therefore means four different things depending on who is charging you. This page collects the official rates for ten families as of August 2026, then normalizes them to the one comparable figure: what a 1024×1024 image costs on each. Every number is first-party unless flagged — reseller price lists contradict each other and are excluded on principle.
Why can't you compare image model prices at face value?
Because four incompatible billing units are in production simultaneously:
- Per image — a flat rate per generation: Google (Imagen 4), Alibaba (Qwen), Ideogram, Recraft, Luma. Simplest to budget — but Luma's $0.0404 buys native 2K while Imagen's $0.04 buys 1K–2K.
- Per megapixel — Flux 2: a 4MP render costs four times a 1MP render at the same tier.
- Per token — GPT Image and Microsoft's MAI Image encode the output image as tokens and bill per million; no flat per-image figure exists.
- Per operation — Reve prices its edit endpoint per edit, independent of what the edit does.
The unit is a tell for the vendor's mental model: per-image sells pictures, per-megapixel sells pixels, per-token sells compute, per-edit sells retouching.
What does each family officially charge?
The per-image group: Imagen, Qwen, Ideogram, Recraft, Luma, and Flux's flat-rate entries
| Family | Tier | Official price | The catch |
|---|---|---|---|
| Imagen 4 | Fast / Standard / Ultra | $0.02 / $0.04 / $0.06 | Shuts down on the Gemini API August 17, 2026 |
| Qwen | 2.0 / 2.0 Pro | $0.035 / $0.075 | 100 free images per model on Alibaba Cloud Model Studio |
| Qwen (editing) | edit-plus / edit / edit-max | $0.03 / $0.045 / $0.075 | The newer editor is the cheapest entry |
| Ideogram 4 | Turbo → Quality | ~$0.03–$0.10 by speed | v3 endpoints confirmed at $0.03/$0.06/$0.09 |
| Recraft | V4.1 / V4.1 Pro | $0.035 / $0.21 | After the June 20, 2026 cuts; V4-era SVG endpoints listed $0.08–$0.30 |
| Luma | uni-1.1 / uni-1-max | $0.0404 / $0.10 | Billed at native 2K; edits and references add ~$0.003 each |
| Luma (legacy) | Photon / Photon Flash | $0.015 / $0.004 | December 2024 models, 2MP output |
| Flux | klein / Kontext pro / Kontext max | $0.014–0.015 / $0.04 / $0.08 | The family's flat-rate entries; the rest is per-MP |
Sources: Gemini API pricing, Alibaba Cloud Model Studio, Recraft's changelog, lumalabs.ai/api, and BFL's rate card, August 2026. Three rows deserve a second look. Imagen 4 Fast at $0.02 is the cheapest first-party image generation Google has ever sold, and it dies with the model on August 17 — the designated successor, Nano Banana, costs $0.039. Qwen's edit-plus at $0.03 undercuts Qwen's own older editor at $0.045 — the better model costing less. And Luma's uni-1.1 rate buys native 2K, quietly one of the best prices per pixel here.
The per-megapixel group: Flux 2
Black Forest Labs' official rates: pro at $0.03/MP for creation and $0.045/MP for edits, flex at $0.05–0.06/MP, max at $0.07/MP. (BFL's own pages disagree between $0.05 and $0.06 for flex — treat it as a range.) The scheme's virtue is transparency: a 1MP draft on pro costs about three cents, a 4MP final about twelve, and you always know why. No other family prices resolution this explicitly.
The token-metered group: GPT Image and MAI Image
OpenAI's official rates per million tokens, August 2026:
| Model | Text in | Image in | Image out |
|---|---|---|---|
| gpt-image-2 | $5 | $8 ($2 cached) | $30 |
| gpt-image-1-mini | $2 | $2.50 | $8 |
(The deprecated gpt-image-1.5 and gpt-image-1 bill $32 and $40 per million image-output tokens until their December 1 and October 23, 2026 shutdowns.) What a single image costs depends on size and quality: roughly half a cent for a small/low GPT Image 2 output up to about $0.21 for large/high — a derived range, not a list price.
Microsoft prices MAI Image the same way: MAI-Image-2.5-Flash at $19.50, 2.5 at $47, and 2.5-Pro at $106 per million image-output tokens (text input $1.75–$5, image input $1.75–$8). No flat per-image figure exists — more pixels, more tokens, end of story.
The per-operation group: Reve
Reve publishes exactly one rock-solid number: $0.01 per edit on Reve Edit Fast, per its Replicate listing. There is no official per-image generation price; third-party hosts quote roughly $0.20–$0.25 for a full-resolution Reve 2.1 generation — directional at best. The penny edit is the story regardless: on Replicate's public registry, Reve's edit endpoints have out-run its create endpoint, 154K+ runs to ~109K.
What does a 1024×1024 image cost on every model?
Normalized to one square-megapixel-class image (per-MP rates ≈ per-image at this size; hedged figures flagged):
| Model | ~Cost per 1024×1024 | Basis |
|---|---|---|
| Luma Photon Flash | $0.004 | Official (Dec 2024 model) |
| GPT Image 2, low/small | ≈$0.005 | Derived from token rates — not a list price |
| Flux 2 klein | $0.014–0.015 | Official |
| Imagen 4 Fast | $0.02 | Official — retires Aug 17, 2026 |
| Flux 2 pro | ≈$0.03 | $0.03/MP official |
| Qwen edit-plus / Ideogram 4 Turbo | $0.03 / ≈$0.03 | Official |
| Krea v2 Medium | ≈$0.03 | Third-party listings only — no official rate card |
| Recraft V4.1 / Qwen 2.0 | $0.035 | Official |
| Imagen 4 Standard / Flux Kontext pro | $0.04 | Official |
| Luma uni-1.1 | $0.0404 | Official — billed at native 2K |
| Flux 2 flex | ≈$0.05–0.06 | Official per-MP (pages disagree) |
| Imagen 4 Ultra / Krea v2 Large | $0.06 / ≈$0.06 | Official / third-party only |
| Flux 2 max | ≈$0.07 | Official per-MP |
| Qwen 2.0 Pro | $0.075 | Official |
| Flux Kontext max | $0.08 | Official |
| Ideogram 4 Quality / Luma uni-1-max | ≈$0.10 | Official |
| GPT Image 2, high/large | ≈$0.21 | Derived from token rates |
| Recraft V4.1 Pro | $0.21 | Official |
| Reve 2.1 generation | ~$0.20–0.25 | Third-party only — directional |
| MAI-Image-2.5 family | No flat figure | $19.50–$106 per M image-output tokens |
The read: $0.03–$0.04 is the market's clearing price for a production-quality frame — Flux pro, Qwen, Ideogram Turbo, Recraft V4.1, Imagen Standard, and Luma uni-1.1 all landed there independently. Below the band you are buying speed; above it, a named capability — grounding web search at $0.07, quality tiers at $0.10, native SVG at $0.21. A model off the band with no capability to point at is simply expensive.
Where do edit prices break from generation prices?
Editing has its own price list, and almost nobody compares it. The spread as of August 2026: Reve at $0.01 per edit; Qwen edit-plus at $0.03; Flux Kontext pro at $0.04; Luma at $0.0434 (a $0.003 premium over its create rate); Flux 2 pro edits at $0.045/MP — 1.5× its own creation rate; Kontext max at $0.08. GPT Image and MAI meter edits at the same token rates as generation — fitting for unified models like Luma's Uni line, "a multimodal reasoning model that can generate pixels" in Luma's words, where an edit is a generation. That 10× spread flips conclusions: a workflow that generates once and edits ten times should be priced off the edit column, where Reve sits in a category of its own.
What people ask about image-model pricing
What is the cheapest AI image model in 2026?
On official rates: Luma Photon Flash at $0.004 per image, though it is a December 2024 model. Among current-generation models, Flux 2 klein at $0.014–0.015 is the cheapest flat rate; GPT Image 2's low tier works out to roughly half a cent, but that is derived, not published.
Why doesn't GPT Image 2 have a per-image price?
OpenAI bills image output in tokens ($30 per million on GPT Image 2), and token count scales with size and quality. Per-image cost is therefore a range — roughly $0.005 to $0.21 — not a number.
Is Imagen 4's $0.02 tier still available?
Only until August 17, 2026, when Google shuts Imagen 4 down on the Gemini API. Nothing first-party replaces the price: the migration target, the Nano Banana line, starts at $0.039.
Are Reve and Krea prices official?
Mostly no. Reve's $0.01-per-edit rate is published on its Replicate listing, but neither vendor publishes a first-party per-image rate card as of August 2026 — the ~$0.20–0.25 (Reve) and ~$0.03–0.06 (Krea) figures circulating come from third-party hosts.
Which image models have real free tiers?
Alibaba gives every Qwen image model a 100-image free quota on Model Studio, and Reve's first-party app includes a free tier. Recraft's free plan has a serious catch: Recraft owns free-plan images, with no commercial license.
Do you have to manage any of this billing yourself?
Not if you run the models through invideo. Every family on this page — token-metered, per-megapixel, per-edit, and flat-rate alike — sits in the agent's image picker among 200+ models, so the four-unit accounting stays on the platform side and model choice reverts to fit for the job. Family depth lives on the hubs linked throughout, plus the AI models index for Krea, MAI, and Luma; the AI image generator workflow is the shortest route from prompt to output.
Version history: first published August 2026, covering official rates across the GPT Image, Imagen 4, Flux 2, Qwen, Ideogram, Recraft, Reve, Krea, MAI, and Luma families.