Why the Same AI Model Costs Different Amounts at Different Providers

You have picked a model. Then you check what it costs and find it listed at $0.03 at one provider, $0.032 at another, and $3 at a third. Same model, same weights, same outputs — three prices, and none of them is a typo.

This is the single most common source of confusion when buying AI generation, and it is also the reason most "cheapest AI model" tables on the web are wrong. They compare numbers that are not measuring the same thing.

Here is what actually varies, with real examples you can check yourself.

Pricing last checked: August 26, 2026. Provider prices change often. The live numbers are on the ElliSekiz compare page.

The short answer

Three things differ between providers, and only one of them is the price:

1. The unit — what one dollar buys. Per image, per megapixel, per second of video, or per second of machine time.
2. The tier — most providers charge more for higher resolution, and most publish only the entry-level number.
3. The rate itself — margin, hardware, and temporary discounts.

Two prices can only be ranked when they land on the same unit. Everything below is a variation on that one rule.

Reason 1: The price is not measured in the same thing

Qwen-Image Edit is one model. Three providers serve it, and all three publish a price that starts with "$0.03".

A megapixel is one million pixels. A 1024×1024 image is about 1.05 megapixels, so fal.ai's per-megapixel rate works out to roughly $0.031 — within a cent of the other two. A 2048×2048 image is about 4.19 megapixels, so the same rate becomes about $0.126.

That is the whole point. At small sizes the three providers are within 7% of each other. At 2K, one of them costs four times what the others do — and its published number never changed. The ranking flipped because the output got bigger, not because anyone raised a price.

So the question "which provider is cheapest for Qwen-Image Edit?" has no answer until you say what you are generating. You can see all three offers side by side on the Qwen-Image Edit comparison page.

When the bigger number is the smaller price

The clearest case in the whole catalogue. The sync/lipsync-v3 model, at the two providers that serve it:

fal.ai's number is thirty-six times larger and its price is lower, because a minute is sixty seconds. Anyone ranking these two by the printed figure would pick the expensive one and believe they had saved money.

Units you will actually run into

• Per image / per generation / per output — one flat price per result. Easiest to budget.
• Per megapixel — scales with resolution. Cheap for thumbnails, expensive for print.
• Per second of output video — scales with clip length.
• Per second of run time — scales with how long the machine was busy. See Reason 3.
• Per million tokens — the text-model unit, with separate input and output rates.

Only prices in the same row of that list can be compared.

Reason 2: Providers tier by resolution, then publish one number

Wan 2.7 Text-to-video is served by all three providers, and all three charge exactly the same headline rate.

At 720p they are identical. Move to 1080p and fal.ai's own page says the rate becomes $0.15 per second, a 50% increase, while Replicate and Atlas Cloud publish a single rate with no resolution tier at all.

This is where a lot of budget surprises come from. The number on the pricing page is almost always the entry tier — the cheapest resolution, the shortest duration, the fastest sampler. If you plan to ship 1080p, the comparison you did at 720p does not hold.

It also means "no tier published" is not the same as "one flat price". It may mean the provider does not vary the price by resolution, or it may mean they have not documented it. Those are different facts, and only the first one is safe to build a budget on. When we cannot tell which is which, we say so rather than guess.

You can run the model at both resolutions from the Wan 2.7 Text-to-video model page and compare the two bills directly.

Reason 3: Some providers bill machine time, not output

This is the biggest one, and it is easy to miss because the numbers look so different that people assume they misread.

Replicate serves Wan 2.2 Animate at $3 per thousand seconds of run time — that is $0.003 for each second the GPU spends working, which is not the same as a second of finished video.

You cannot turn that into a price per video. Run time depends on the length of the input clip, the resolution, the number of steps, and which GPU picks up the job. A 5-second output might take 40 seconds of run time or 400. The provider is not hiding anything; they genuinely do not know in advance, so they bill for what they can measure.

The practical consequences:

• Per-output pricing is predictable, run-time pricing is not. If you need to quote a customer a fixed price per generation, per-output billing is worth paying a premium for.
• Run-time pricing can be much cheaper at scale, especially for short or simple jobs, because you stop paying for the provider's safety margin.
• A run-time rate can never be ranked against a per-output rate. Anyone who puts $0.003 next to $0.10 and calls the first one cheaper is comparing a machine-second to a video-second.

On ElliSekiz these rows are shown but deliberately not ranked, and the row says the comparison could not be made. A missing ranking is information; a fabricated one is not.

Reason 4: Some of the price is a temporary discount

Providers run promotions, and the promotional rate is often the only one you see.

Atlas Cloud's own model catalogue publishes both a list price and a current price. Today, three examples:

If you build a cost model on the discounted number, your unit economics have a hidden expiry date. It is worth knowing which of your prices are promotional and re-checking them before you commit to volume.

A caveat on the Atlas figures: the catalogue publishes a number per model without an explicit unit label. Where other providers serve the same model, the unit can be corroborated — the video models above are per second of output, matching fal.ai and Replicate for the same models. Where no other provider serves it, the unit is inferred, and we flag it rather than present it as verified.

Reason 5: Sometimes it really is just a different price

After you strip out units, tiers, run time and discounts, real differences remain — and they can be large.

Happy Horse 1.1 Text-to-video, priced by both providers in the same unit, per second of output video:

Same model, same unit, exactly double the price. This is a genuine comparison and Atlas Cloud is genuinely cheaper for this workload today.

The reasons are ordinary business reasons: different hardware, different utilisation, different margin, different negotiated rates with the model's publisher. None of that is visible from the outside, which is exactly why the number has to be checked rather than assumed.

You can generate the same clip at both providers from the Happy Horse 1.1 model page and compare the output as well as the price.

How to compare two prices without fooling yourself

A short procedure that works for any pair of providers:

1. Write down the unit before the number. Not "$0.03" but "$0.03 per megapixel". If you cannot name the unit, you do not have a price yet.
2. Convert to cost per finished thing. One image at your real resolution. One clip at your real length. Conversion is only legitimate when it is arithmetic — $0.50 per 5 seconds really is $0.10 per second. Turning a run-time rate into a per-image price is not arithmetic, it is a guess.
3. Check the tier. Is the published number the entry resolution? What do you actually ship?
4. Check whether it is a promotion. If the provider publishes a list price, note it.
5. Only then compare. If the units still do not match after step 2, the honest answer is that these two prices cannot be ranked.

If two providers land within a few percent of each other, price is not your deciding factor. Latency, availability, rate limits and output quality are — and those you have to test.

Which provider should you choose?

• You ship at high resolution. Avoid per-megapixel billing. A flat per-image price that looks slightly more expensive at 1MP will usually win at 2K and above.
• You need predictable unit costs. Choose per-output billing, even at a premium. Run-time billing makes per-customer pricing very hard to quote.
• You run large volumes of short, simple jobs. Run-time billing is worth modelling — it can be substantially cheaper, but only measure it with your own workload.
• You are still prototyping. Pick on quality and availability, not price. At prototype volumes the difference between $0.030 and $0.032 is not worth a day of your time.
• You are committing to volume. Re-check every number, note which are promotional, and ask the provider about committed-use rates.

Checking today's numbers

Provider prices move, promotions expire, and new models arrive weekly. Any table — including the ones in this article — is a snapshot.

ElliSekiz keeps a price ledger across fal.ai, Replicate, Atlas Cloud and OpenRouter that records each price in the unit that provider bills in, with the date it was verified, and refuses to rank two prices whose units do not match. That is the same discipline described above, applied to the whole catalogue instead of one model at a time.

The compare page shows every model and provider in one table, and each model has its own page — for example Qwen-Image Edit or Wan 2.7 Text-to-video — with the current offers side by side.

Conclusion

The same model costs different amounts at different providers for five reasons, in roughly this order of impact:

1. The billing unit is not the same thing.
2. The published number is the entry resolution tier.
3. Some providers bill machine time rather than finished output.
4. Part of the price may be a temporary discount.
5. Real rate differences, which can be 2× or more on identical units.

Get the unit right and most of the confusion disappears. What is left is a real comparison, and often a real saving.

Explore more on ElliSekiz

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

Then test the model. Every model page has a playground, so you can run the same prompt at more than one provider and compare the output as well as the bill.

Sources

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

• fal.ai — sync lipsync v3
• fal.ai — Qwen-Image Edit
• fal.ai — Wan 2.7 Text-to-video
• fal.ai — Happy Horse 1.1 Text-to-video
• Replicate — Qwen-Image Edit
• Replicate — Wan 2.7 Text-to-video
• Replicate — Wan 2.2 Animate
• Atlas Cloud model catalogue API, https://api.atlascloud.ai/api/v1/models