How Much Does AI Actually Cost? Three Worked Examples

"How much does AI cost?" has no single answer, and the reason is not that providers are being cagey. It is that three different products are being priced in three different units, and the thing that moves your bill is usually not the price on the pricing page.

So instead of a price list, here is a method and three budgets worked end to end: a product-image pipeline, a chat feature, and a batch of short videos. Every number is checked, every assumption is stated, and you can swap your own numbers into the same arithmetic.

Pricing last checked: August 26, 2026. Prices move. The live figures are on the ElliSekiz compare page.

The method

Four steps, in this order. Skipping step 1 is what produces wrong budgets.

1. Name the unit. Per image, per megapixel, per second of output video, per second of machine time, per million tokens. A number without a unit is not a price.
2. Convert to cost per finished thing. One image at the resolution you actually ship. One clip at the length you actually need. One conversation, not one token.
3. Multiply by real volume, including the work you throw away.
4. Then compare providers. Not before.

The examples below all follow it.

Example 1: 10,000 product images

The task. An e-commerce catalogue needs 10,000 generated product shots, text-to-image, 1024×1024, one output per prompt.

So the answer is $270 to $380 — a 40% spread for the same job. Worth ten minutes of comparison, not worth a week.

Three things are worth noticing.

The cheapest number is a promotion. Atlas Cloud's $0.027 for Seedream 4 is discounted from a $0.03 list price. Budget on the list price and treat the discount as upside.

Compound prices need unpacking, and rounding matters. Replicate charges FLUX.2 pro as $0.015 per run plus $0.015 per output megapixel plus $0.015 per input megapixel — about $0.030 for a one-megapixel text-to-image with no input image.

fal.ai prices the same model as "$0.03 for the first megapixel of output, plus $0.015 per extra megapixel of input and output, rounded up to the nearest megapixel". That rounding is the catch. A 1024×1024 image is 1.05 megapixels — a hair over one — so it rounds to two, and the same job costs $0.045 rather than $0.030. Fifty percent more, for four percent more pixels.

Two formulas that look equivalent, and are not, because of one word about rounding.

Resolution changes the ranking, not just the total. Move to 2048×2048 and Atlas Cloud's flat per-generation price stays at $0.03, while a per-megapixel price multiplies by roughly four. The flat-price provider was slightly more expensive at 1MP and is dramatically cheaper at 2K. If you do not know your output resolution yet, you do not have a budget yet.

You can see the current offers for these models side by side — Seedream 4, Wan 2.7 Text-to-image — and run the same prompt through each from its model page.

Example 2: a chat feature, 50,000 messages a month

Text models are priced per million tokens, with separate rates for input and output. A token is roughly a short word or part of one; the exact count depends on the model's tokenizer.

The assumptions. Stating them is the entire job here — change any one and the answer changes.

• 50,000 messages per month
• 1,200 input tokens per message — a system prompt, a few turns of history, and the user's question
• 300 output tokens per message — a short, useful answer
• Which gives 60M input tokens and 15M output tokens per month

$9 to $675 for the same feature — a 75× spread, far wider than anything in the image example. Model choice dominates here in a way it does not in media generation.

Output tokens dominate, even though there are fewer of them

There are four times as many input tokens as output tokens in this workload, and output still costs as much or more than input on every model in the table. That is because output is priced 4× to 8× higher per token than input.

The practical consequence: the cheapest reliable way to cut a text bill is to make the answers shorter. Trimming average output from 300 to 200 tokens saves a third of the larger half of the bill. On Claude Sonnet 5 that is $50 a month for one line of prompt instruction.

The system prompt is charged on every message

If 800 of those 1,200 input tokens are a system prompt you wrote once, you are sending it 50,000 times a month — 40M tokens, or $80 a month on Sonnet 5, for text that never changes.

Prompt caching exists exactly for this: the provider stores the unchanging prefix and charges a fraction to re-read it. If your system prompt is long and your traffic is steady, this is usually the single biggest saving available, and it requires no model change.

Watch for introductory pricing

Claude Sonnet 5's $2 / $10 is an introductory rate that runs through August 31, 2026. The standard rate is $3 / $15, which turns this same workload from $270 into $405 a month — a 50% increase on a date already in the calendar.

This is not unusual and it is not a trick; new models often launch with promotional pricing. It is simply something a twelve-month budget has to account for.

Token prices for every chat model on the platform are on the chat models page.

Example 3: 200 short marketing videos

The task. 200 clips, 5 seconds each, 720p, text-to-video. That is 1,000 seconds of output.

$60 to $420 — a 7× spread, and note that two rows of it are the same model at different providers.

The lever that matters most is not in the table. Clip length multiplies everything. Ten-second clips instead of five doubles every number above. So does deciding you need two takes per concept. Video is the category where a small change in the brief moves the budget more than any provider decision.

Resolution does the same. fal.ai's Wan 2.7 rate is $0.10 per second at 720p and $0.15 at 1080p — the same 200 clips become $150 instead of $100.

Current rates are on the Kling 2.5 Turbo Pro comparison page, and you can generate a test clip from its model page before committing to 200.

Putting it together

A small product doing all three things in one month:

Both columns are real, defensible budgets for the same feature set. The difference is roughly 4×, and most of it is the chat model.

Five things that move your bill more than the provider you pick

1. Output resolution and clip length. The single largest multiplier in media generation, and the one most often left undecided until after the budget is written.
2. Output token length. In text, this is where the money is. Shorter answers are cheaper answers.
3. The system prompt. Charged on every request. Cache it.
4. Work you throw away. Rejected generations, retries after a failure, and A/B variants all cost full price. Whatever your reject rate turns out to be, it applies as a multiplier to every number above — so measure it during your pilot rather than assuming it away.
5. Promotions expiring. Note which of your prices are introductory, and when they end.

What these numbers do not include

Honesty about scope matters more than a bigger total:

• Storage and bandwidth for the media you generate
• Any moderation, upscaling or post-processing step
• Your own infrastructure and engineering time
• Failed and rejected generations, per the point above
• Minimum commitments or platform fees, where a provider charges them

Model inference is usually the largest line in an AI feature's budget, but it is rarely the only one.

Checking today's numbers

Every figure here is a snapshot taken on August 26, 2026. Providers change prices, promotions end, and new models arrive weekly.

ElliSekiz tracks prices across fal.ai, Replicate, Atlas Cloud and OpenRouter, recording each one in the unit that provider bills in and stamping it with the date it was verified. The compare page is the live version of every table above.

Conclusion

There is no single price for AI, but there is a reliable method:

• Name the unit before the number
• Convert to cost per finished thing at your real resolution, length and answer size
• Multiply by real volume, including the work you discard
• Compare providers last

Run that on your own workload and you will get a number you can defend — usually somewhere between the two columns above, and almost always more sensitive to your product decisions than to which provider you sign up with.

Explore more on ElliSekiz

Compare before you commit. Every model and provider in one table, each price in its own billing unit with the date it was checked — on the compare page.

Test with your own workload. Every model page has a playground, so you can measure real output lengths and real reject rates before you multiply anything by 10,000.

Sources

All prices were read from the providers' own sources on August 26, 2026.

• Atlas Cloud model catalogue API, https://api.atlascloud.ai/api/v1/models
• Replicate — Seedream 4
• Replicate — Wan 2.7 Image
• Replicate — FLUX.2 pro
• Replicate — Kling 2.5 Turbo Pro
• fal.ai — FLUX.2 pro
• fal.ai — Wan 2.7 Text-to-video
• OpenRouter models API, https://openrouter.ai/api/v1/models
• Anthropic pricing