The Cost of AI Images: Credits, Compute, and When Stock Is Cheaper

AI image generation feels free. You type a prompt, an image appears, and it costs you — what, exactly? A fraction of a cent in API credits? One of your monthly generations? The vague sense that you're "using up" something? The actual cost of AI-generated images is harder to calculate than it should be, and the honest math doesn't always favor the AI tool over the stock photo. This article does the math.

The headline finding: AI images are dramatically cheaper than custom photography or illustration, roughly comparable to stock photography for simple use cases, and potentially more expensive than stock for anyone who values their time honestly. That last part is the one nobody talks about, because the per-image cost hides the per-usable-image cost, and those are very different numbers.

What It Actually Does

Let's start with the direct costs — what you pay to generate images across the major platforms as of early 2026.

Midjourney charges $10/month for the Basic plan (roughly 200 generations), $30/month for Standard (15 hours of GPU time, which works out to roughly 900 generations in relaxed mode), and $60/month for Pro (30 hours fast, unlimited relaxed) [VERIFY]. The per-image cost on the Standard plan works out to roughly $0.03-0.05 per generation. On the Basic plan, it's closer to $0.05. These numbers assume you use most of your allocation each month — if you generate 50 images on a $30 plan, your effective per-image cost is $0.60.

DALL-E / GPT Image through ChatGPT Plus costs $20/month as part of the subscription, with a generation limit that varies by model and usage [VERIFY]. Through the API, GPT Image (gpt-image-1) costs roughly $0.02-0.19 per image depending on resolution and quality settings [VERIFY]. The API pricing is straightforward but the subscription pricing is bundled — you're paying for the LLM too, so attributing a per-image cost is fuzzy. If you're already paying for ChatGPT Plus for the text features, the image generation is effectively a bonus. If you're subscribing primarily for images, the math is less favorable.

Flux through the API (via Replicate, fal.ai, or BFL's own API) costs roughly $0.01-0.06 per image depending on the model tier and provider [VERIFY]. Flux Pro through BFL's API is at the higher end. Flux Schnell — the fast, lower-quality model — is at the lower end. Running Flux locally on your own GPU costs nothing per image but requires hardware. A capable consumer GPU (RTX 4090) runs Flux Dev at reasonable speeds, and the "cost" is the $1,600+ for the card plus electricity. If you generate thousands of images per month, local runs cheaper. If you generate dozens, the API is cheaper.

Stable Diffusion locally is the budget option. The models are free. The cost is your hardware and your time. A reasonable local setup — an RTX 3090 or 4070 Ti with 16GB VRAM — generates SD XL images in under 10 seconds [VERIFY]. The per-image cost is effectively just electricity, measured in fractions of a cent. But the setup cost is real: the GPU, the ComfyUI or Automatic1111 configuration, the model downloading and testing, the workflow building. For someone who already has the hardware and enjoys the tinkering, this is the cheapest option by far. For everyone else, the setup time is the cost — and it's significant.

Adobe Firefly is included with Creative Cloud subscriptions and available standalone at $5/month for 100 generative credits [VERIFY]. The per-image cost is roughly $0.05 on the standalone plan. The value proposition isn't price — it's the cleaner training data provenance and IP indemnification on enterprise plans.

Stock photography, for comparison, costs roughly $0.20-10.00 per image depending on the service and subscription tier. Shutterstock's standard subscription runs about $29/month for 10 images ($2.90 each) to $199/month for 350 images ($0.57 each) [VERIFY]. Adobe Stock is similar. Unsplash and Pexels are free for most uses, though the selection for specific needs is limited. Custom photography runs $200-2,000+ per session. Custom illustration runs $50-500+ per piece.

The raw per-image cost comparison favors AI generation by a wide margin — sometimes by a factor of 10x or more. A Midjourney image costs $0.03-0.05. A stock image costs $0.57-2.90. A custom illustration costs $50+. The math seems obvious. It's not.

What The Demo Makes You Think

The per-image cost is a lie. Not because the numbers are wrong, but because they measure the wrong thing.

The number that matters is the per-usable-image cost — the total cost divided by the number of images you actually use. And this is where AI image generation gets expensive, because the hit rate is low. Not zero. But low.

Here's what a typical AI image generation session looks like in practice. You need a hero image for a blog post about automation workflows. You write a prompt. The result is close but the composition is wrong. You revise the prompt. Now the composition is better but the style doesn't match your brand. You revise again. The style is closer but there's an artifact — a weird hand, a text element that's gibberish, a background element that doesn't make sense. You revise again. You try a different seed. You try a different model. Forty minutes and twelve generations later, you have an image you can use.

Those twelve generations cost you $0.36-0.60 in credits. They cost you 40 minutes of time. If your time is worth $50/hour — a conservative number for a professional — that 40 minutes cost you $33. Your "cheap" AI image actually cost $33.60. The stock photo that would have taken two minutes to find and download cost you $2.90 plus $1.67 in time — $4.57 total.

This is not a contrived example. I've tracked my own AI image generation across several months, and the average time-to-usable-image for anything beyond a simple, generic illustration is 15-30 minutes. For images with specific requirements — matching brand colors, consistent character design, precise compositions, text-in-image — it's often longer. The time cost dominates the credit cost by an order of magnitude.

The demos never show this. They show the prompt that worked — the one-shot generation that produced exactly the right image. They don't show the eleven prompts that preceded it. They don't show the 40-minute prompt engineering session. They definitely don't show the moment when you gave up, opened Shutterstock, found the right image in 90 seconds, and spent the remaining 38 minutes doing actual work.

There's another hidden cost: the compute and infrastructure for teams. If you're running an agency or content operation that generates hundreds of images per month, the API costs add up in a way that individual subscription plans don't. A content team generating 1,000 images per month through the Midjourney API or DALL-E API might spend $50-200 in direct costs [VERIFY] — still cheap compared to stock or custom. But the team time spent prompting, reviewing, revising, and quality-checking those images is the real expense. At scale, the question isn't "is AI cheaper per image" but "is the total cost of the AI workflow — tooling, prompting, QA, revision — cheaper than the total cost of the stock workflow — search, license, download."

What's Coming (And Whether To Wait)

Three trends will change the cost calculation over the next year.

Quality is improving, which reduces the revision cost. The hit rate — the percentage of first-attempt generations that are usable — has improved meaningfully from 2024 to 2026. DALL-E 3 and GPT Image handle composition, text rendering, and consistency significantly better than DALL-E 2 did. Midjourney v6 produces more reliable results than v5. As models get better, the number of revisions per usable image drops, and the time cost drops with it. The trajectory here is clear: the per-usable-image cost is falling, and it's falling because of quality improvements, not pricing changes.

Consistency tools are emerging. One of the highest time costs in AI image generation is maintaining visual consistency — generating images that share the same style, color palette, and visual language across a project. Character references, style references, and fine-tuning are getting better. Midjourney's style reference and character reference features, Flux's LoRA training, and DALL-E's reference image support all reduce the time spent wrestling consistency out of models that naturally tend toward variation. When these tools mature — and they're close — the per-usable-image cost for branded content drops significantly.

Pricing is compressing. The cost per API call has dropped substantially from 2024 to 2026, and the trend will continue as compute gets cheaper and competition intensifies. The direct cost of generation is converging toward trivial for standard resolutions. The price war benefits users, but it also means the direct cost is becoming a smaller fraction of the total cost. The bottleneck is increasingly time, not credits.

Should you wait for cheaper or better before investing in AI image generation? No — but you should be realistic about what you're investing. The tool subscriptions are cheap. The time to learn effective prompting, build consistent workflows, and develop a reliable quality bar is not cheap. That investment pays off over months, not days. Start now, but don't expect the per-usable-image cost to match the per-generation cost until you've built real fluency with your chosen tool.

The Verdict

The honest cost framework for AI images in 2026 has three tiers.

AI wins clearly when you need high volume, generic imagery — blog post illustrations, social media graphics, placeholder visuals, internal presentations. The per-image cost is pennies, the quality is good enough, and the time cost is manageable because the specificity requirements are low. Stock photography can't compete on volume economics, and custom work can't compete on either volume or cost.

Stock wins when you need a specific, professional image quickly and it exists in the stock library. The search-and-download workflow takes two minutes. The AI prompt-and-revise workflow takes twenty. For the professional who values their time, the stock image at $2.90 is cheaper than the AI image at $0.05 plus 20 minutes of prompting. This is especially true for common subjects — business settings, nature scenes, food, architecture — where stock libraries have millions of high-quality options.

Custom work wins when you need something distinctive, brand-specific, or legally clean for high-stakes commercial use. A professional illustrator or photographer produces exactly what you need, with full copyright, in a predictable timeframe. The cost is higher — dramatically higher per image — but the total project cost might not be, because you don't spend hours prompting and revising.

The mistake is treating AI image generation as universally cheaper. It's universally cheaper per generation. It's not universally cheaper per usable output when you factor in time. The right approach in 2026 is to use all three — AI for volume and experimentation, stock for speed and reliability, custom for quality and distinctiveness — and choose based on the actual need, not the per-credit pricing that makes AI feel like the obvious answer for everything. It's a good tool. It's not always the cheap tool, despite what the pricing page suggests.

Updated March 2026. This article is part of the Image Generation series at CustomClanker.

Related reading: AI Images vs. Stock Photos, AI Image Ethics and Copyright, Midjourney vs. DALL-E vs. Flux