Where 5p comes from

The number

About 5p a kit, on top of a run that costs 13p once research is cached

How it's known

Every call writes its tokens and cost next to what it produced

What you get

A cost per customer you can state before anyone signs up

A real moment: pricing day

Sooner or later a feature built on a language model needs a price. That day goes very differently depending on what was recorded along the way.

Where the money goes in a research run

In the research stage, over 90% of the cost is input tokens. The reason is simple once you see it: every web page the model fetches stays in the conversation, and is sent again on each round that follows.

So a generous page size isn't paid for once. It's paid for on every later round. That single observation is where most of the saving came from.

Four settings, about 70% less

Four configuration changes brought a research run down by roughly 70%. No prompt rewrite, no new architecture — four numbers.

The reasoning behind the model choice is worth copying. Research is extraction: read pages, pull out facts, cite the source. A mid-tier model does that well. Choosing which opportunity is worth an organisation's effort is judgement, so that step stayed on the strongest model, where it costs about 11 to 16 US cents a call.

Research is then cached per organisation for 30 days. The first run pays for the web research, measured at £0.51 to £1.16. Every run in the following 30 days reuses it and costs about 13p.

Why the kit is only 5p

The kit applied those lessons from day one.

1. 1No web tools. The research already exists. The kit call receives up to twelve verified facts from it and nothing else, so there's no growing conversation to pay for.
2. 2One call. The story, the levels, the email, four social posts and the plan messages come back in a single structured response. Measured cost: 6 US cents, about 5p, in about 25 seconds.
3. 3Edits cost nothing. Changing a giving level updates the story, the email and the buttons with no model call at all. Part 3 shows how.
4. 4Rewrites are capped. Each idea gets its first kit plus three rewrites.
5. 5Every call is costed. Token counts and cost are written to the run at the rates for the model that actually ran, so the figure in the database matches the invoice and stays right when the model changes.

Caps turn a variable cost into a known one

Two runs a month. Four kit generations per idea. Customers rarely notice limits like these, and they do something valuable: they put a ceiling on what any one customer can cost.

For a subscriber, before payment fees:

The ceiling assumes every limit is reached, which almost nobody does. That's the point of it: the price can be set against a number that cannot be exceeded, rather than against a hope.

The same logic prices the free trial. A trial is one run, so welcoming a new organisation costs at most about £1.20 in model spend. A free trial with a known price is a marketing budget. One without is an open tab.

A newer model is only an upgrade if you measure it

When two newer models arrived, both were run against the stored research before anything changed. One was on par with the current choice and slightly cheaper. The other, at the same effort, returned about half as many verified facts.

Nothing was switched. The test took an afternoon, cost pennies, and replaced a guess with a result. Stored research makes that kind of comparison nearly free, which is one more reason to keep it.

What to take from this part

• Measure before you price. The pricing only exists because every run records its own cost.
• Caps turn a variable cost into a known one. They are product decisions with a finance benefit.
• Give a free trial a known price. One run, one ceiling.
• Spend the model on judgement. Dates, sums, formatting and validation are cheaper, faster and more reliable in ordinary code.
• Sell the finished job. Nobody is buying "AI". They're buying a launch they can run this afternoon.