How much does it cost to generate an AI image? We tested 19 models
In 2024 you could spot an AI image by the broken text. All 19 models in this test wrote our poster text correctly, so what separates a one-cent image from a 23-cent one today is brief-following, style and layout. We ran the same three briefs on every model: a photo, a poster and a flat illustration.
An AI image generator comparison is only useful if every model gets the same brief, so that is what we did: one photography prompt and one text-heavy poster prompt through all 19 image models in the catalog we use for ad production, plus a flat-illustration prompt through the catalog's nine illustration-capable models, one take each, no cherry-picking. The whole test cost $2.62. Prices in this article range from $0.01 to $0.23 per image, and all 19 models wrote the poster text correctly.
That last sentence is the big finding. Two years ago, "can it write text?" was the first question you asked of an image model. Today the interesting failures are somewhere else entirely: small text in colors you can't read, briefs quietly ignored, and two separate models printing the hex code from my prompt onto the image as if it were copy.
The short version
- Price spread is 23x, quality spread isn't. The $0.01 renders from FLUX 3 Klein and Recraft V4.1 Flash were usable; the $0.23 Recraft V4.1 Pro render was better, not 23 times better.
- Big text is solved. 19 out of 19 models wrote "LAUNCH DAY" and the five-word line under it correctly. The failures moved to layout, color and size.
- Prompt leakage is the new typo. Krea 2 Large printed the hex code "0144F" on the poster; Seedream 4.5 painted "#014F5" on a coffee mug. Exact color codes in prompts are a known trap.
- Style obedience varies more than quality. On the flat-vector brief, some models stayed flat, others drifted into 3D renders or added effects nobody asked for.
- For ads, pick per job, not per leaderboard. Photography, flat illustration and typography have different winners, and the arena-topping model costs 6x the one a few places below it.
How we tested 19 models
The method is the same one behind our 24-model video test and comes from our AI ads cost project, where every example ships with its exact cost. Three prompts, every model, one take each, identical settings per run (3:4 aspect, 1K resolution where the model offers it). Prices are what the catalog charged per job; it sells generation at cost, so they sit close to the makers' list rates, with gaps where resolution tiers and billing units differ. No regenerates, no "best of four". What the model returned first is what you see. Each image is captioned with the price we were actually charged and the wall-clock render time.
- The photo: a founder's desk at dusk, laptop dashboard, electric blue mug, specific light directions. Tests photorealism and art direction.
- The flat illustration: a three-color vector scene with a person, three robots and a staircase that rises like a chart. Tests palette discipline and style obedience. Nine models ran this one, the catalog's illustration-capable set: both FLUX 3s, all three Recraft tiers, Seedream 5.0 Flash, Nano Banana 2.1, Qwen Image 3 Pro and Grok Imagine 2.0.
- The poster: a launch poster with big title text, a smaller line under it and an isometric rocket. Tests text at two sizes plus layout.
The lineup is every image model in one production catalog: FLUX 3 Image and Klein from Black Forest Labs, Google's Nano Banana 2.1, ByteDance's Seedream 4.5 and 5.0 Flash and Lite, Recraft V4.1 in Flash, base and Pro, Alibaba's Qwen Image 2.1 Turbo and Pro plus Qwen Image 3 and 3 Pro, Krea 2 in Medium Turbo, Medium and Large, Sourceful's Riverflow 2.5 Fast and 2.6 Pro, and xAI's Grok Imagine Image 2.0. Two absences worth naming up front: Midjourney has no API and DALL-E no longer exists, covered below.
Every model, price and render time
Prices are what we were charged per image (identical across the three runs); render time is from the poster run. The one-line verdict is from the poster test, the hardest of the three briefs.
| Model | Maker | Per image | Render | Poster verdict |
|---|---|---|---|---|
| Recraft V4.1 Flash | Recraft | $0.01 | 4 s | Nice linework; pink sub-line nearly invisible |
| FLUX 3 Image Klein | Black Forest Labs | $0.01 | 8 s | Clean type and layout at a penny |
| Qwen Image 2.1 Turbo | Alibaba | $0.02 | 12 s | Crisp text, but went 3D instead of flat |
| Seedream 5.0 Flash | ByteDance | $0.02 | 15 s | Clean lines plus tasteful flourishes |
| Krea 2 Medium Turbo | Krea | $0.02 | 16 s | Headline fine; sub-line washed out |
| Riverflow V2.5 Fast | Sourceful | $0.03 | 24 s | Balanced layout, both lines clean |
| Qwen Image 3 | Alibaba | $0.04 | 76 s | Added a drop shadow nobody asked for |
| Recraft V4.1 | Recraft | $0.04 | 5 s | Sub-line exiled to the bottom edge |
| Krea 2 Medium | Krea | $0.04 | 25 s | Clean, with a pseudo-3D phone |
| Seedream 5.0 Lite | ByteDance | $0.04 | 30 s | Clean type; phone drifted photoreal |
| Nano Banana 2.1 | $0.05 | 13 s | Faithful flat style, dynamic layout | |
| Qwen Image 3 Pro | Alibaba | $0.05 | 74 s | Flat and friendly, on brief |
| Qwen Image 2.1 Pro | Alibaba | $0.05 | 15 s | Sharp type; glossy 3D, not flat |
| Seedream 4.5 | ByteDance | $0.05 | 8 s | Big bold text, clean concept |
| FLUX 3 Image | Black Forest Labs | $0.06 | 20 s | On-brief flat style, crisp at both sizes |
| Grok Imagine Image 2.0 | xAI | $0.07 | 70 s | On-brief, clean type at both sizes |
| Krea 2 Large | Krea | $0.07 | 55 s | Missed the brief; printed the hex code |
| Riverflow V2.6 Pro | Sourceful | $0.14 | 111 s | Polished, but reinvented the layout |
| Recraft V4.1 Pro | Recraft | $0.23 | 8 s | Poster-grade type; rocket became an outline |
Two patterns in the numbers. Recraft is the speed outlier in both directions: its Flash tier renders in 2 to 4 seconds, and even its $0.23 Pro tier comes back in 9 or under, while Qwen Image 3 and Grok take over a minute. And the price column splits the catalog into a sub-nickel iteration tier and a two-digit-cents finishing tier, which is exactly how I'd use them: explore cheap, finish expensive.
All 47 images, side by side
Three briefs, every model, one take each. Switch between the photo, the illustration and the poster. Each caption carries the model, what it charged, the render time and what stood out. You can also download the full dataset behind each run (illustration, poster).
The prompt: A founder's desk at dusk, photographed from a low three-quarter angle. A laptop open on a simple analytics dashboard with one line rising to the right, a ceramic mug in electric blue (#0144F5) in the foreground, warm lamp light from the left, soft window light behind, shallow depth of field, calm. No people, no readable text, no logos.
The poster test: text rendering is (mostly) solved
All 19 models wrote both lines of the poster correctly. "LAUNCH DAY" in large type, "Now on the App Store" in small type, not a single wrong letter across the whole set. If you last checked in 2024, when broken text was the giveaway of an AI image, that result alone is worth updating on. Google leaned hard into text with its Nano Banana Pro launch, which claimed "state-of-the-art text rendering"², and the rest of the field has visibly caught up on big text.
What's not solved is everything around the letters:
- Small type gets sacrificed. Recraft V4.1 Flash and Krea 2 Medium Turbo both rendered the sub-line in a pale pink that nearly vanishes into the background; Recraft V4.1 shrank it and dropped it at the bottom edge of the frame. The words are correct, the hierarchy is wrong.
- Prompt values leak into the art. My prompts specified colors as hex codes, and two models treated them as copy: Krea 2 Large set "0144F" next to an Apple logo on the poster, and Seedream 4.5 printed "#014F5" across the mug in the photo test. Both also misquoted the code, dropping a digit. If you prompt with hex values, check the output for them.
- Style instructions are negotiable. "Clean vector style" lost to a glossy 3D render on both Qwen 2.1 variants, and Qwen Image 3 added a drop shadow to the title on its own. On the illustration brief, FLUX 3 Klein quietly introduced a fourth color into a three-color palette.
The QA question for AI statics has changed. Stop checking the big text for typos; start checking the small text contrast, the palette, and whether your own prompt is printed somewhere in the image.
One more result I didn't expect: on the illustration brief, Nano Banana 2.1 read "a staircase that rises like a chart" and drew the staircase as actual bar-chart columns. That's the kind of interpretation that makes an illustration feel designed rather than generated, and no other model tried it.
What an AI image costs in 2026
List prices, from the makers' own pricing pages this week: FLUX 3 Image costs $0.048 at 1K resolution and its Klein sibling $0.009¹. Nano Banana 2.1 lists at $0.0336 at 1K, half of what its predecessor charged (our catalog billed it at $0.05)². Recraft V4.1 runs from $0.007 for Flash to $0.21 for Pro's 2K output, with SVG vector output at $0.08³. Qwen Image 2.1 Turbo is $0.016 per image on Alibaba Cloud⁴, and Grok Imagine Image 2.0 is $0.04 on xAI's API⁷.
Watch the resolution cliffs, though. FLUX 3 Image jumps from $0.10 at 2K to $0.607 at 4K, while Klein does 4K for $0.036¹. If you need big canvases, say for print or heavily cropped multi-format ad sets, the model you iterate with and the model you upscale with shouldn't automatically be the same one.
For scale: a human designer averages around $48 an hour in the US, and a one-off poster typically quotes at $100 to $500¹⁰. The entire 47-image test in this article cost $2.62, which is why the economics of ad statics have stopped being about generation cost at all. The cost is now in the taste: knowing which 3 of 47 images are usable, and fixing the sub-line before it ships. That matches where the industry already is: 86% of ad buyers told the IAB they're using or planning to use generative AI for ad creative, and it expects AI involvement in video creative alone to reach 39% by 2026⁸.
Where Midjourney and DALL-E went
The two names most people would type into this comparison aren't in it, for opposite reasons.
Midjourney never shipped an API. As of late 2026 it remains subscription-only, $10 to $120 a month through the web app and Discord, currently on V8⁶. It produces beautiful images and you cannot build a production pipeline on it: the third-party "Midjourney APIs" are Discord-automation wrappers that violate its terms. For a one-off brand shoot, fine. For programmatic ad creative, it's disqualified by architecture, not quality.
DALL-E is gone. OpenAI shut down API access to DALL-E 2 and 3 on May 12, 2026, pointing everyone at GPT Image⁵. GPT Image 2.6 currently tops the Artificial Analysis blind-preference arena with an Elo of 1259, but it costs about $0.21 per image. Nano Banana 2.1 sits in the cluster just below it at $0.034⁹. That 6x price gap for a visible-but-not-6x quality gap is the whole story of image model pricing right now.
Which model for which job
Based on this test and the thousands of statics we generate for ad accounts each month, here's how I'd assign them. Disagree after looking at the galleries; that's what they're for.
| Job | First pick | Cheap alternative | Why |
|---|---|---|---|
| Product/lifestyle photo | FLUX 3 Image ($0.06) | FLUX 3 Klein ($0.01) | Most natural light handling in the photo test; Klein gets 90% of the way |
| Flat illustration | Recraft V4.1 Pro ($0.23) | Recraft V4.1 Flash ($0.01) | Cleanest figure work; Flash keeps the palette at 1/23rd the price |
| Type-led poster | FLUX 3 Image ($0.06) | Grok Imagine 2.0 ($0.07) | Both held hierarchy at two text sizes without inventing effects |
| High-volume iteration | Recraft V4.1 Flash ($0.01) | Seedream 5.0 Flash ($0.02) | 2 to 15 second renders; cheap enough to generate 50 and pick 3 |
| Literal brief-following | Grok Imagine 2.0 ($0.07) | Nano Banana 2.1 ($0.05) | The two that followed all three briefs most exactly |
The judgment call I'll defend: don't pay arena-leader prices for ad statics. Blind-preference rankings reward beauty on arbitrary prompts, and ads don't need the most beautiful image. They need the right image: on-palette, on-layout, correct at a glance, shipped in volume and refreshed before fatigue. In this test the gap between $0.05 and $0.23 was taste, not capability, and the brief-following failures happened at every price point. The production question this test doesn't answer is consistency, whether image 50 can match image 1's style, which is a reference-and-brand-kit problem more than a model problem. That's the next test.
Frequently asked questions
What is the best AI image generator in 2026?
On blind human preference, OpenAI's GPT Image 2.6 leads the Artificial Analysis arena, but at roughly $0.21 an image it costs six times more than Nano Banana 2.1, which sits a few places below it at $0.034. For ad work, our test suggests the honest answer is a pair: Nano Banana 2.1 or FLUX 3 Image for photography, and Recraft V4.1 or FLUX 3 for flat illustration. The quality gap between a one-cent and a twenty-cent image is now smaller than the style gap between any two models.
Can AI image generators render text correctly?
For big text, yes. All 19 models in our test wrote a two-word title and a five-word line under it correctly, which would not have happened in 2024. The failures that remain are smaller: small text in washed-out colors, drop shadows nobody asked for, and two models that printed the hex color code from the prompt onto the image as visible text.
How much does an AI-generated image cost?
Between about $0.007 and $0.23 per image through current APIs. In our 19-model test the cheapest passes were Recraft V4.1 Flash and FLUX 3 Image Klein at about a cent each, and the most expensive was Recraft V4.1 Pro at $0.23 for a 2K render. All 47 images in this article together cost $2.62 to generate.
Why isn't Midjourney in this comparison?
Midjourney still has no official public API as of late 2026. It is subscription-only at $10 to $120 a month through its web app and Discord, currently on the V8 model line. The third-party "Midjourney APIs" you can find are unofficial wrappers around Discord automation, and using them risks account suspension, so no serious ad workflow should be built on one.
Is DALL-E still available?
No. OpenAI shut down API access to DALL-E 2 and DALL-E 3 on May 12, 2026 and points developers at the GPT Image models instead. GPT Image is strong (it tops the blind-preference arenas) but it is billed in tokens and works out to roughly $0.21 per high-quality image, which is why cheaper catalog models cover most ad use cases.
Do AI-generated images work in paid ads?
They already run everywhere: the IAB projected that 39% of video ad creative would be made or enhanced with generative AI by 2026, and 86% of ad buyers were using or planning to use it. Meta applies a detection-based "AI info" label but has no disclosure mandate for ordinary commercial ads. The practical constraint is not policy, it is quality control: the hex-code leaks in our test are exactly the kind of thing you have to catch before an image ships.
Sources
- FLUX 3 Image and Klein per-image API pricing by resolution — Black Forest Labs, API pricing (2026)
- Nano Banana 2.1 pricing and Google's text-rendering positioning — Gemini API pricing (2026) and Engadget on Nano Banana Pro (2025)
- Recraft V4.1 tier pricing and vector output — Recraft API docs (2026)
- Qwen Image 2.1 Turbo and Pro API pricing — Alibaba Cloud blog (2026)
- DALL-E 2 and 3 API shutdown on May 12, 2026 — OpenAI deprecations (2026)
- Midjourney API status and subscription pricing — MyArchitectAI, Midjourney API overview (2026)
- Grok Imagine Image 2.0 API pricing — xAI model docs (2026)
- Generative AI share of ad creative, 2024 to 2026 — eMarketer on the IAB Digital Video report (2025)
- Blind-preference Elo and per-1,000-image prices — Artificial Analysis text-to-image arena (as rendered Nov 8, 2026)
- Freelance designer rates and poster pricing — PayScale (2025) and Thervo graphic design cost guide (2025)
- Meta's AI-content labeling approach — Meta newsroom (2024, updated 2025)