Three cents an image sounds like nothing. Four thousand eight hundred generations later it is $144, and you have 1,200 usable images to show for it — which means your real cost per keeper was twelve cents, four times the number on the price card. Nobody budgets for the rejects, and the rejects are most of the bill. What follows is what actually drives the figure this calculator returns.
1. The retry multiplier decides the answer
Set the model to FLUX.2 Pro at $0.030 an image and ask for 1,200 usable images a month. At an acceptance rate of one in four:
1,200 × 4 = 4,800 generations × $0.030 = $144 a month, or $0.12 per usable image.
Now switch the model to Kling Image 2.1 at $0.014 — less than half the price — and hold the retry rate at four. You land at $67.20. You saved $76.80. Then leave the model alone and improve your prompting and reference workflow until acceptance moves from one in four to one in two: 1,200 × 2 × $0.030 = $72. Almost the same saving, and you kept the model whose output you preferred.
That is the general shape of this market. The spread between the cheapest and dearest credible image model is roughly five and a half times — Kling Image 2.1 and FLUX.2 Klein 4B at $0.014 against Stable Image Ultra at $0.080 — while the spread between a sloppy pipeline and a disciplined one is easily ten. Fix the multiplier first.
2. What acceptance rates actually look like
Guessing this field is the fastest way to a useless answer, so count it for one week. Log every generation and every asset that reached final use. Until you have that number, these are the ranges studios report:
| Brief type | Generations per keeper |
|---|---|
| Loose stock-style imagery, no brand constraint | 1.5 – 2 |
| Blog and social illustration, defined style | 3 – 4 |
| Product imagery matching a real SKU | 6 – 10 |
| Character or brand consistency across a set | 8 – 15 |
| Text rendered accurately inside the image | 10+ |
The pattern is that constraint costs money. A brief with three hard requirements — this product, this colour, this composition — multiplies attempts rather than adding to them, because each requirement independently gives the model a way to miss. If your acceptance rate sits above six, the fix is a reference-image or editing workflow, not a cheaper model.
3. Video compounds because it bills by the second
Image pricing is per unit. Video pricing is per second of output, and the duration multiplies against the retry count, which is where budgets detonate.
A 30-second clip on Veo 3.1 at $0.40 a second costs $12 per attempt. Not per finished clip — per attempt. Three attempts to get one usable 30-second clip is $36. Ten such clips a month is $360, before a single revision requested by a client.
The calculator's default is 300 finished seconds a month at three generations per usable clip, which is 900 billed seconds. Across the index that is:
- Veo 3.1 Lite at $0.05/s — $45
- Veo 3.1 Fast at $0.10/s — $90
- Veo 3.1 at $0.40/s — $360
- Veo 3.1 4K at $0.60/s — $540
- Kling 2.5 Turbo at $0.042/s — $37.80
- Luma Ray3.2 at $0.06/s (720p) — $54
- FLUX 3 Video HD at $0.17/s — $153
Same brief, same output length, a fourteen-fold spread. This is why the calculator prices every model at your volume rather than just the one you selected: the ranking rarely surprises anyone, but the size of the gaps almost always does.
4. Resolution is a pricing tier, not a setting
Luma Ray3.2 costs $0.06 a second at 720p and $0.24 a second at 1080p. Identical model, identical prompt, four times the price for more pixels. On our 900 billed seconds that is $54 against $216. Veo 3.1 shows the same structure more gently: $0.40 a second at 720p/1080p, $0.60 at 4K, a 50% uplift.
The workflow that follows is obvious and widely ignored. Iterate at the cheap tier until the shot is right, then render the final take once at the expensive one. Applied to the Luma example — twelve exploratory 10-second attempts at 720p ($7.20) plus one 10-second final at 1080p ($2.40) — you spend $9.60 where iterating entirely at 1080p would have cost $31.20. The saving is 69% and the finished asset is pixel-identical.
The same logic applies across model families, not just within them. Block out at Veo 3.1 Lite, finish at Veo 3.1.
5. When a subscription beats the API
Per-generation API pricing is linear: every asset costs the same as the last one. Subscription tools are the opposite — a fixed monthly fee, a capped or throttled allowance, and a marginal cost of zero until you hit the ceiling.
Midjourney Standard is $30 a month for 15 hours of fast GPU time plus unlimited relaxed generation. Against 4,800 FLUX.2 Pro generations at $144, the subscription is cheaper by a wide margin, and the unlimited relaxed queue means a patient team can run far past that. Runway Standard is $12 a month for 625 credits and Pro $28 for 2,250; Pika Standard is $28 for 700.
The subscription wins on three conditions, and loses if any of them fails. First, a human is driving — subscriptions come with an interface, not an endpoint, so anything that needs to fire from your own code is out. Second, your volume sits inside the allowance, or you can tolerate the slow queue. Third, the licensing terms suit your use; check them, because they vary and commercial rights are not uniform across tiers.
Cross those out and you are on the API, where the price is the price and the retry multiplier is the only lever you control.
Then check what you are not counting
Storage, CDN delivery, moderation passes, upscaling, and the human hours spent reviewing 4,800 images to pick 1,200 — at ten seconds an image that is thirteen hours a month of somebody's time, which on any realistic loaded rate exceeds the $144 you spent generating them. That is the honest total, and it is the strongest argument for spending a week improving prompts. Our hidden costs of AI guide covers the rest of that ledger, the full price index lists every model here with its source, and if you are weighing a subscription against metered access more broadly, the subscription versus API calculator runs the same comparison for text.
Prices used by this calculator were last verified on 15 August 2026 from the vendors' own pricing pages. See the full price index for every figure and its source, the change log for what has moved recently, and the methodology for how the index is maintained and where its limits are.