What a credit actually is, and the day we measured it

Most usage-based AI pricing is an allowance with a marketing name on it. You are sold units, or actions, or points, and nowhere on the page can you work out what the thing you actually do will cost. The number is real. It is just not convertible.

A credit here is a unit. One million tokens of the base model is 100,000 credits, and heavier models cost proportionally more from the same balance. That is the exchange rate, and it is the one the billing code uses.

The rest of this is the half almost nobody publishes: what one run of each feature cost, on a day we can name, with the command that produced it and the arithmetic to check it.

An allowance you cannot convert is not a price

The definition is one sentence, and the pricing page carries it word for word: a credit is a real unit, not an allowance, and you can see every run and what it cost in your account. Two things in it do the work. One currency, so a heavier model draws harder on the same balance rather than emptying a second wallet you did not know you had. And a published rate, so a balance is a quantity of model work rather than a quantity of goodwill.

An allowance is not dishonest by itself. What makes it useless is that you cannot turn it into work you recognise. Knowing the rate is not enough either: nobody signing up knows how many tokens a call analysis sends. Publishing that is our job rather than the reader's.

The day we measured it

We ran the measurement on 2026-08-16. One command, manage.py measure_feature_credits, against a representative discovery call of about twenty-five minutes. What it produced are single runs of the real code path, not medians. There is nothing to take a median of: the platform holds no customer AI usage, so calling any of it an average would be a claim about data that does not exist. The pricing page says so under the table.

One call analysis cost 8,526 credits on that run, at the pinned smart model. One cycle of live call coaching cost 1,159, at the fast one. The four call-preparation features each came in under a thousand, and one question asked of the copilot mid-call came in a little over, all on the same day and out of the same command.

A single run is more representative than it sounds, for a reason in the code rather than in the fixture. Both hot paths are capped: the analysis sends the last 12,000 characters of transcript plus a brief capped at 2,500 characters, and the live coach sends a rolling window of twenty transcript segments plus 1,200 characters of brief. Any call long enough to fill those windows costs what a much longer one costs. What varies is how much the model writes back, and that is the smaller half of the total.

The basis is the part worth publishing

Every published figure carries a basis, and it takes one of two values. Measured means what it says: one run of the real code path, on the date beside it, at the model named in the row.

Structural zero is the more interesting one. Recording and transcription cost nothing, and neither does voice mode, and that is not a measurement at all. The engine is local whisper, its row in the model price table is priced at zero, and the multiplier derived from that row is therefore zero. There is no run to count tokens from because there is no bill to count. Nothing about it could come out differently next time, which is why the note beside it says plainly: free, and not as an introductory offer.

Publishing the basis matters more than publishing the number. A bare number can only be believed or not. A number with its basis attached can be argued with: your calls are shorter than the fixture, or you would run the cheaper model, or the date is old enough to ask about.

That is also why the pricing configurator labels its estimates as estimates. Three numbers there are neither measured nor typed by you: about two coaching suggestions per ten minutes of call, about one question asked of the copilot per thirty minutes while the call runs, and about one asked before it. All three are printed next to the controls they qualify, so you know where to substitute your own number.

The arithmetic, so the number can be re-checked

The sum is published beside the figures, so a credit cost can be checked without running a model. Add the billable tokens, multiply by the model multiplier, divide by ten, round up.

Take the call analysis run. It reported two input tokens, 6,793 output tokens, 3,039 read from cache and 7,218 written to it. Add those four, multiply by five, which is the smart model's multiplier, divide by ten and round up, and you get 8,526. The coaching cycle is the same sum at the baseline multiplier, and it is where the rounding up is visible: its tokens divide to a fraction, and the charge published is 1,159.

Two things fall out of that arithmetic. Cache tokens are billed at full rate by design, which is why they dominate: on the analysis run, more tokens were written to and read from cache than the model wrote in its answer. And multipliers are not typed into a file. They are derived from the model price table, which refreshes hourly, and the rule the table is kept to is written down beside the figures: if a model's price moves, re-run the command and republish.

That is what the date on the table is for. The figures are pinned to two models we name rather than describe: claude-haiku-4-5, the baseline, and claude-opus-5, whose multiplier is five. Change either and every figure moves.

Every run is a line in your own account

None of this asks you to take the table on faith. Every run is a line in your credits breakdown, grouped by the feature that spent it, so if a published figure were wrong for your workspace, your own breakdown is the thing that would say so.

That itemisation is what makes an unattended spender something a person leaves switched on. Autopilot is off until you turn it on, and switched off it costs nothing. Turned on, its round bills under its own key, so a night of drafting is a row of its own rather than part of one undifferentiated total. The limit worth knowing: work the round hands to another part of the product, a task analysis for instance, lands on that part's line instead.

The same holds for anything you build yourself. API and MCP calls are priced rather than measured, on Business and Professional only: 10 credits per call, a flat published rate, itemised like every other run. A call that fails is not billed at all, because a refusal is not service delivered and billing it would make a retry loop a way to drain your own balance. A loop that keeps succeeding is the other case: it spends exactly as a person clicking would, and the breakdown is where you find that out.

What one run does not tell you

A single run is not a promise about your month, and the pricing page does not pretend otherwise. Your own calls will vary, mostly in how much the model writes back and in how often the live coach speaks up, which depends on the conversation rather than on a clock.

The configurator that multiplies these figures stops rather than extrapolates. The seat rail and the calls-a-month rail each have a top stop, and past either one the page stops quoting a price. One measured run per feature is a fair basis for an ordinary month, not for a month far larger than the one measured.

If you would rather check this against your own conversation than ours, start on the free plan. It opens every screen, carries 10,000 credits a month and takes no card, and on the figure above that is about one analysed call. What it does not carry is the work that runs without you, live guidance included. Record a real call, let it analyse, then read your own credits breakdown. It is the only measurement in this article that is about you.

Questions this raises

Why one measurement instead of an average?

Because there is no usage history to average. The platform holds no customer AI usage, so one run of the real code path is what we can publish instead. It is labelled as one, dated 2026-08-16, and it ships with the token counts and the model multiplier so you can reproduce the figure without running anything.

What happens when I run out of credits?

The AI work stops; the workspace does not. Your meetings, clients, products, invoices, tasks and files are all still there and still usable. Add a top-up, which never expires and lands instantly, or wait for the monthly allowance to reset.

Do API and MCP calls cost credits?

Yes, on the plans that have them: API keys and the MCP server are on Business and Professional. It is 10 credits per call, a flat published rate rather than a measured run, drawn from the same balance as everything else and itemised in your account the same way.