What Do You Actually Pay Per AI Image in 2026?
Between roughly US$0.005 and US$0.21 per image, depending on the model, the quality tier, and the resolution. But that range hides the real problem: the tools you are choosing between do not bill the same way. Some sell you GPU time, some sell you tokens, and only one of those lets you calculate a per-image price at all.
That mismatch is why the "which AI image tool should I pay for" question stays confusing longer than it should. You are not comparing prices. You are comparing pricing models.
The extractable numbers, as of 3 August 2026:
--- Midjourney bills by fast GPU time, not by image. Basic US$10/mo (3.3 fast GPU hours), Standard US$30/mo (15 hours plus unlimited Relax mode), Pro US$60/mo (30 hours), Mega US$120/mo (60 hours). Annual billing takes 20% off, bringing them to US$8, US$24, US$48 and US$96.
--- Nano Banana 2 (Google, launched February 2026) is priced per output: about US$0.067 at 1K, US$0.101 at 2K, US$0.151 at 4K.
--- FLUX.2 Pro sits at roughly US$0.03 per image on fal.ai, with the hosted pro, flex and max tiers landing in the US$0.03 to US$0.07 range.
--- GPT Image 2 is token-billed, at US$30 per million image output tokens plus US$8 per million image input tokens. Per-image estimates commonly quoted are roughly US$0.005 at Low, US$0.041 at Medium, and US$0.165 at High.
In Hong Kong dollars at HK$7.8 to the US dollar, that means about HK$0.23 per FLUX.2 Pro image, HK$0.52 per 1K Nano Banana 2 image, and HK$1.29 per high-quality GPT Image 2 image. Midjourney has no equivalent number, and the next section explains why.
Why Does Midjourney's US$10 Plan Have No Per-Image Price?
Because Midjourney sells you fast GPU hours, not images. The Basic plan's 3.3 fast GPU hours works out to roughly 200 images if you use the allowance efficiently, which is where the widely quoted figure of about US$0.05 per image comes from. Generate slower, more complex, or higher-resolution work and that number moves.
This matters more than it sounds. A subscription with a GPU-time budget rewards planning and punishes iteration. Every reroll spends the same budget as the original, so the "generate twenty and pick one" habit that produces good creative work is exactly the habit that burns the allowance.
The Standard plan at US$30 is where most regular users land, and the reason is not the 15 fast hours. It is unlimited Relax mode, which lets you queue images slowly at no additional GPU cost. If your work can tolerate a wait, Relax mode is the cheapest high-volume image generation on the market, because your marginal cost per image drops toward zero.
Two more plan facts worth knowing before you pick a tier. Pro at US$60 exists mainly for Stealth Mode, which keeps your generations out of the public gallery, and that is a client-confidentiality feature, not a quality feature. Mega at US$120 is aimed at studios doing high-volume video work.
There is also no free trial. Basic at US$10, or about HK$78 a month, is the cheapest way in.
How Much Do the API-Billed Models Cost Per Image?
Between about US$0.03 and US$0.17 for a single standard-resolution image, with no monthly commitment. API-billed models charge per generation, so a month where you make 12 images costs roughly a twentieth of a month where you make 240. That is the structural advantage over any subscription.
The three that matter for practitioner work in August 2026:
--- FLUX.2 Pro at roughly US$0.03 per image is the cheapest of the three, and reporting places its photorealism close to Nano Banana Pro at a lower price point. HK$0.23 per image.
--- Nano Banana 2 at US$0.067 for 1K scales cleanly: US$0.101 at 2K and US$0.151 at 4K. You pay for resolution, which means you should stop asking for 4K on drafts. HK$0.52, HK$0.79 and HK$1.18 respectively.
--- GPT Image 2 is the one to watch on billing mechanics. It is token-based, not per-image, and the cost you pay depends on two things you control directly: the quality tier and the output resolution. Low, Medium and High are roughly HK$0.04, HK$0.32 and HK$1.29.
Run the arithmetic on a realistic month. Two hundred images at FLUX.2 Pro is about US$6, or HK$47. The same 200 images on GPT Image 2 at High quality is about US$33, or HK$257. Midjourney Basic covers roughly the same 200 images for US$10, HK$78, with no per-call setup at all.
That comparison is the actual decision. Not which model is best, but which billing model matches your volume.
Which Option Is Right for How Many Images You Make Per Month?
Under 50 images a month, use API-billed models and pay only for what you make. Between 50 and 400, a Midjourney subscription is usually cheaper and simpler. Above 400, Midjourney Standard with unlimited Relax mode is difficult to beat on cost per image.
The verdict by user type:
--- Occasional user, under 50 images a month: API-billed FLUX.2 Pro at about HK$0.23 per image. Fifty images is HK$12. Any subscription is worse arithmetic at this volume.
--- Solo marketer or content creator, 100 to 300 images a month: Midjourney Basic at HK$78, or Standard at HK$234 if you need Relax mode and reroll freely. Best aesthetic quality per dollar in this band.
--- High-volume producer, 400+ images a month: Midjourney Standard at HK$234, using Relax mode for anything not urgent. Marginal cost per image approaches zero.
--- Client work requiring confidentiality: Midjourney Pro at HK$468, because Stealth Mode is the only reason that tier exists. Do not pay it for image quality.
--- Anyone building an automated pipeline: API-billed models, because a subscription cannot be called programmatically from a workflow. FLUX.2 Pro for volume, GPT Image 2 High for the few images that need to be right.
One pattern shows up repeatedly in practitioner reporting and it is worth taking seriously: most marketing teams do not use a single image generator for everything. They pair two or three across different stages, cheap and fast for drafts, expensive and precise for the final asset. That is not indecision, it is correct cost routing, and it is the same logic that applies when routing video generation across models or when choosing which text model should be your default.
What Do These Prices Not Include?
The generation fee is the smallest line in your real cost. Every option on this list has limitations that either add cost or rule it out for specific work, and the honest version of this comparison names them rather than hiding them in a footnote.
What each option gets wrong:
--- Midjourney: No free trial, so you cannot evaluate before paying. GPU-time billing means heavy iteration eats your allowance unpredictably, and the per-image cost you actually pay is only knowable in hindsight. There is no API-first workflow, so it does not slot into an automated pipeline.
--- Nano Banana 2: Resolution-tiered pricing means costs escalate quietly. A team that defaults to 4K pays roughly 2.25 times the 1K rate for images that get resized down for web use anyway.
--- FLUX.2 Pro: Cheapest per image, but you are buying through a hosting provider, which introduces a dependency on that provider's uptime, rate limits, and pricing changes rather than a direct vendor relationship.
--- GPT Image 2: Token-based billing makes budgeting genuinely hard. The commonly quoted per-image figures are calculator estimates derived from token rates, not list prices, so your actual invoice depends on prompt length and output resolution in ways that are not obvious in advance.
--- All of them: The generation fee excludes your time. At 30 minutes of prompting and selection per usable image, a HK$0.23 image with a HK$300/hour opportunity cost is a HK$150 image. Tool price is rarely the dominant term.
Licensing is the other thing to check before you commit, not after. Commercial-use rights vary by plan and by provider, and it is a plan-level detail on some tools rather than a blanket policy.
And to be equally direct about our own position: UD does not sell an image generation model, and we publish no fixed price for image-generation work, because the cost depends entirely on volume, review workflow, and how much of the process gets automated. Any number we quoted here without seeing your workflow would be invented. If your bottleneck is purely "which subscription do I buy", the answer is above and it does not involve us.
How Do You Test Cost Before Committing?
Generate the same brief on two models at the cheapest tier, count how many attempts each needed to produce one usable image, then multiply. Cost per usable image is the only number that matters, and it is always higher than cost per generation because most generations are not usable.
Use this prompt to build the test brief so both models get identical instructions:
Try this prompt:
--- I need to run a cost comparison between two AI image models for [describe your actual recurring image need, for example: product shots on plain backgrounds for an e-commerce listing].
--- Write me one single image prompt, 40 to 60 words, that is specific enough to produce a usable result on the first or second attempt. Include subject, framing, lighting, background, and style. Do not include any model-specific syntax or parameters, because I need to run the identical prompt on two different models.
--- Then list the 5 specific things I should check in the output to decide whether that image is usable or needs a reroll, phrased as pass or fail checks, not as opinions.
--- Finally, give me a 3-column table to record: attempt number, pass or fail, and which of the 5 checks failed.
Run that prompt on 10 images per model, at the cheapest quality tier, and record the results. If Model A needs 1.4 attempts per usable image at HK$0.23 and Model B needs 3.1 attempts at HK$0.23, Model A is less than half the cost even though the sticker price is identical.
Ten images at FLUX.2 Pro rates is about HK$2.30. That is the entire cost of making this decision with evidence instead of a blog post, including ours.
The Right Next Step
If you are choosing a tool this week, run the 10-image test above before you subscribe to anything. It costs a few Hong Kong dollars and it replaces every ranked list on the internet with a number that describes your work.
If the tool is already chosen and the problem is that the workflow around it is slow, that is a different problem and no subscription fixes it. The cost that hurts is the review, resizing, filing and re-briefing that happens after generation, and that is the part worth automating.
We understand AI. We understand you better. With UD by your side, AI doesn't feel cold. Twenty-eight years spent helping Hong Kong teams work out which part of the problem is actually the expensive one.
See Where Your AI Spend Is Actually Going
Picking a model is the easy part. Building the workflow around it, so generation, review and publishing run without you babysitting each step, is where the time goes. We'll walk you through every step, from comparing models on your own briefs to designing and deploying the workflow.
Reviewed by the UD AI team. Prices were checked on 3 August 2026 against vendor pricing pages and published pricing reporting. AI image pricing changes frequently and per-image figures for token-billed models are estimates, not list prices. Verify current rates before committing budget.
