AI companies explain what they charge for. Not how you pay.

Half the category never tells you how billing works

Billing framing does not fail gently. Companies either handle it properly or omit it entirely, and the failure is the same one every time: a single cadence on the page, no toggle, no annual saving stated.

• GitHub Copilot shows only a monthly cadence at $10 USD per user / month, omitting annual billing where software buyers expect both.
• Grammarly shows a single cadence where buyers expect two.
• Weaviate offers Pay-as-you-go, monthly, no commitment. on its mid tier with no annual option and no toggle.
• Sourcegraph states Starting at $16K without labelling whether that is annual or monthly.

The companies that got it right did something unremarkable and specific. Otter.ai writes Monthly Annual save 51%. Replit writes Save 20% $25 $20 per month billed annually. Jasper writes Monthly Yearly Save ~20%. Krisp writes Monthly Annually Save 50%.

Steal this: show both cadences with the saving stated as a percentage. It is an afternoon of work, and it is the single change most likely to move a page's score.

The companies selling AI have the worst pricing pages in AI

Developer-facing pages, meaning LLM APIs, inference platforms, vector databases and coding tools, average 68.8. Business-facing AI products average 74.3.

The interesting part is where that gap does not come from. On value-metric clarity, developer-facing pages score 7.5 against 7.4 — they are marginally better at explaining what the unit is. Per million tokens. Per GPU hour. Per second of compute. The clarity is there.

The gap is almost entirely billing framing, at 3.8 against 6.7, with trial and freemium logic second at 6.4 against 7.3.

Five developer platforms score 2 out of 10 on trial logic, with no free credit, trial period or freemium tier stated anywhere on the page: Replicate, OpenRouter, RunPod, Together AI and DeepInfra. These are products whose entire adoption motion is a developer trying something in an afternoon.

They tell you the price of the thing and nothing about the transaction.

One page in sixty-eight uses an anchor

Price anchoring averages 6.12, and the distribution is the finding. Only three pages score 4 or below, so this is not widespread failure. It is widespread adequacy. Almost everyone does the same acceptable thing: mark one plan most popular, and stop.

Exactly one page scores above 8.

Lindy places a decoy at the top of its ladder: Human assistant The classic option. $8,000 / month, sitting beside its AI tiers. Every price below it reads as a saving rather than a cost. The comparison is doing the selling.

Steal this: put the thing your product replaces on the page, with its real cost attached. It is free to add, and in a sample of 68 companies, one had thought of it.

Credits are the new default unit

46 of 68 pages, or 68%, price in credits. Credits have quietly become the default unit of AI pricing, and the trade they carry is measurable.

Against pages that do not mention credits, credit-priced pages score +0.5 on value-metric clarity and -0.6 on billing framing, -0.6 on trial and freemium logic, and -0.7 on tier fencing.

Credits make the unit clearer and the purchase murkier. A buyer can see what a credit buys and cannot see what a month costs, what happens when the credits run out, or how the tiers differ once you strip the credit allowance away.

Steal the craft, skip the privileges

Steal: both billing cadences on the page with the saving as a percentage. An anchor, meaning the thing you replace with its real cost. A stated free allowance in real units rather than the words free trial alone. An explicit statement of what happens when someone hits the limit, which almost nobody writes.

Skip: hiding every price behind contact sales, which large brands can afford and you cannot. Credit units nobody can convert into a number of things they will actually receive. A pricing page written for the engineer who integrates the product rather than the person who approves the spend. Copying a competitor's structure, since half of them are scoring 60 or below too.

Sources

Scores were produced by the Kinetic teardown engine on 5 August 2026. Every figure in this report traces to that run. Company names and quoted page text are reproduced as they appeared on the run date.