What Kie.ai Actually Costs: Credits, Per-Model Rates, and the Fine Print

Kie.ai runs on credits at a published standard rate of $0.005 per credit, with a 5 to 10 percent bonus on larger top-ups. New accounts get 80 free credits, which works out to about forty cents of testing. Credits do not expire, and failed tasks are not charged.

That is the whole billing model, and it is a good one. The complications are elsewhere: the documentation publishes no prices at all, the marketing surfaces disagree with each other about what several models cost, and the savings claims move depending on which page you read.

What the credit rules actually guarantee

Three commitments are worth having in writing, because plenty of platforms offer none of them.

Credits do not expire. Unused balance stays in the account indefinitely. Compare that with the aggregators and studios that reset an allowance every 30 days, where unspent money simply evaporates.

One caveat on that promise, since this article is about giving a company money. The terms state credits never expire, and separately reserve the right to suspend or terminate an account at sole discretion, without notice or liability. No clause anywhere says what happens to an unspent balance when that occurs. Those two provisions have never been reconciled in writing, so top up in amounts you would be relaxed about losing rather than buying a year ahead to chase the bonus.

Failed tasks are not charged. Kie describes this as zero-risk billing, and it matters more than it sounds. Generation failures are routine at volume, and platforms that bill for them turn a flaky model into a line item.

Eighty free credits on signup, no card mentioned. At the standard rate that is roughly $0.40, which is enough to generate a handful of images or a few seconds of video. It is a genuine trial rather than a trap, but it is small, so plan to top up before you can evaluate anything seriously.

Set against that, one limit to design around: generated media is retained for 14 days, logs and metadata for two months, and download URLs expire 20 minutes after they are issued. Kie is a generation service, not storage.

What the models cost

The documentation contains no dollar figures anywhere. These come from the price comparison on the homepage, read on 22 September 2026. The right-hand column is Kie’s own claim about what the upstream provider charges, not a figure I verified with those providers.

The Grok row is the one I can independently sanity-check, because xAI publishes its own rate. It charges five cents a second for video, which matches the official column exactly. That is a point in favour of the rest of the table being honest.

What 80 free credits actually buys

Worth doing the arithmetic, because “80 free credits” is meaningless until you convert it.

At the standard rate those credits represent about $0.40. Against the prices above that is roughly ten images on Nano Banana 2 at 1K, or around thirteen on GPT Image 2, or about thirty seconds of Grok Imagine video at 480p. It does not cover a single Veo 3.1 Quality clip at any of the three prices Kie quotes for one.

So the free tier is sized for checking that your integration works, not for evaluating output quality across models. If you are comparing video models, budget a real top-up before you start, because you will exhaust the free balance on your second or third clip.

One thing that makes this less painful than it sounds: because failures are not billed and credits do not expire, a small top-up genuinely lasts until you spend it. There is no clock running against you while you decide.

Where the numbers stop agreeing

Veo 3.1 carries three different prices across Kie’s own surfaces on the same day:

• the homepage comparison table says $1.28 for a Quality 1080p video
• a featured model card on the same page says $0.025 per request
• the app’s own marketing copy says $0.4 for a Fast video and $2 for a Quality video

These cannot all be right, and no documentation page adjudicates. If Veo is your main workload, generate one clip and read the actual credit deduction before you build a budget on any published figure.

The savings claims wander similarly. The homepage hero says average savings of up to 84 percent. A section further down the same page says many models are around 30 percent lower with selected ones at 60 to 70 percent. The getting-started documentation says typically 30 to 50 percent, reaching up to 80 percent for some models. All four are Kie’s own characterisation of competitors’ pricing.

The honest way to read this: the discount is real and it is large on some models, but treat any single headline percentage as marketing rather than a quote.

When going direct is the better deal

The aggregator discount has to cover the cost of adding a vendor between you and the model. Three cases where it usually does not.

When the official price is already trivial. Grok Imagine at five cents a second means a 15-second clip costs 75 cents direct. Saving three quarters of that is real, but it is saving 55 cents per clip, which is unlikely to justify a dependency unless you are generating at volume.

When you have committed spend. If your team already negotiated a rate or holds a subscription with the provider, routing through a reseller means paying twice for the same leverage: once for the access you already have, and once to the intermediary whose rate now applies instead.

When continuity matters more than margin. A reseller’s catalogue is contingent on agreements you are not party to. Kie has removed at least two model families it once documented: Midjourney, whose API specification survives only in web archives, and Sora 2, whose pages went from live to 404 within hours on 15 May 2026. Neither removal was announced, and the terms contain no clause about upstream provider changes. That is the shape of the risk rather than a criticism: models can leave quietly, and your integration goes with them.

Where the discount genuinely earns its keep is the opposite situation. You are testing many models, you have no committed spend anywhere, and the alternative is opening eight vendor accounts to compare eight video models. One key, one balance that does not expire, and no charge for the failures is a good way to run that comparison.

The practical answer

Fund the account with a small top-up, generate one of each thing you actually intend to make, and read the real credit deductions. That gives you a price list specific to your workload, current as of the day you ran it, which is more than any published table here can offer given the inconsistencies above.

Then decide per model rather than per platform. Nothing stops you using Kie for the long tail you are still evaluating and going direct for the one model that ends up carrying your production traffic.

A note on what you are actually buying

It is worth separating two things that the pricing conversation tends to blur.

The first is the discount, which is straightforward arithmetic you can check per model, and which appears to be real.

The second is the aggregation itself: one key, one balance, one integration shape across a hundred models. For some teams that is worth more than the discount, because the alternative is maintaining eight separate vendor relationships with eight billing arrangements and eight different task formats. For others it is worth nothing, because they use one model and already have an account with its maker.

Kie is unusually direct about the trade in its own documentation, stating that overall stability may be slightly lower than official providers and describing that as a conscious trade-off. Taking a vendor at its word when it volunteers a weakness is reasonable. Price it in: build retries, expect the occasional failed task, and keep the integration thin enough that swapping one model for a direct connection later is an afternoon rather than a rewrite.