The Five-Hour Test

01The moment the meter stops

That prompt is the most honest question in the artificial intelligence industry, and almost nobody is treating it as a measurement. Every other signal we use to judge these companies is downstream of somebody’s intention. A run-rate is a bookkeeping convention applied to contracts a procurement department signed. A valuation is a bet on 2030. A benchmark score is a number produced by the company being scored, on a test that is public while the training is not. A weekly active user is a person who opened a tab.

The continue prompt is different. It arrives unannounced, in the middle of something, when the alternative is genuinely free, and it asks one person to convert their private estimate of value into a dollar amount within a few seconds. That is not a survey. That is a price.

So here is the proposal, and it is a narrow one: if you want to verify that Anthropic is creating real value in people’s lives rather than merely capturing budget, track the dollars that individual subscribers voluntarily spend to continue past a usage limit — month over month. Not seats. Not committed enterprise spend. Not the plan fee. The top-up bought at the wall.

A subscription tells you somebody hoped this would be useful. A top-up at the limit tells you it was.

02Why the scoreboard is contaminated

None of the standard numbers are lies. They are simply answers to questions other than the one I am asking. Each of them can be moved by someone whose incentive is to move it, which is precisely what a good measurement should not permit.

Figure 01 What each headline number actually measures

Metric	What it actually measures	How it moves without value moving
Annualized run-rate	One recent period multiplied out to twelve months. Reported above $65 billion for July 2026.	Multi-year commitments, prepaid balances, and credits can all be recognized into the period being annualized.
Valuation	What a small number of investors believe about the end of the decade.	A single round can reprice the company without a single user changing behavior.
Benchmark scores	Performance on published tests, reported by the entity being tested.	The evaluation is public; the training data is not. Improvement and contamination look identical from outside.
Weekly active users	Curiosity, free tiers, employer mandates, and habit.	A person who opens the product and closes it counts the same as one who finishes something.
Enterprise share of revenue	What procurement believes. Last stated publicly as roughly 80% from business customers, in October 2025.	Seats are bought in blocks and can go unopened for a full contract year.
Voluntary top-ups at the limit	One person deciding, under time pressure, that the next hour of work is worth more than the money.	Hard to inflate — but see section 05. Anthropic has already built two features that blur it.

03The metric, stated precisely

Loose versions of this idea are worthless, because everything interesting hides in the qualifiers. Here is the version I would actually want audited:

Gross dollars of extra usage consumed by individual, non-seat subscribers, purchased after a limit notification, at undiscounted standard rates, with a manual confirmation — reported monthly.

1. Consumed, not purchased

   A balance sitting unspent in an account is a deposit, not a verdict. Only usage that was actually burned counts, and only in the month it was burned.

2. Individual, not seat-based

   When a company pays, the person spending has not made a trade-off with their own money. The whole point of the measurement is that the decider and the payer are the same human being.

3. After the notification

   The purchase has to be a response to the wall, not a plan made in advance. Timing is the information. Buy it a week early and you have expressed optimism; buy it at 4:40 p.m. and you have expressed value.

4. Manually confirmed, undiscounted

   A standing auto-reload rule is a decision made once and then delegated. A volume discount is a bet on future consumption. Both are perfectly reasonable products. Neither is a measurement.

04Free waiting is what makes it clean

Here is the property that makes this signal unusual, and I have not found another consumer market with it. The competing offer to the top-up is the same product, a few hours later, at no cost. Every paid Claude plan runs on a rolling five-hour session window, with weekly caps stacked on top. When you hit the wall, the identical capability becomes available again, for free, on a clock.

Compare that with almost anything else people buy under pressure. A cancelled flight has no free version of itself four hours away. Neither does a surgeon, a lawyer, or a plumber. In nearly every purchase we use to reason about willingness to pay, the alternative to buying is going without. Here, the alternative is going without briefly, and the buyer knows exactly how briefly, because the interface tells them.

Two more details make the signal cleaner than I expected when I went looking.

• Extra usage is opt-in and off by default
• Priced at standard API rates, no markup
• Explicit consent required each time
• Prepaid balance, not a postpaid bill
• Daily redemption capped at $2,000

You have to go into settings and turn this on before it can ever be offered to you. It is not a trap that springs; it is a toll booth you had to install yourself. And because the rate is the plain published API rate rather than a penalty price, nobody buying it can be said to have been gouged into it. Whatever those dollars represent, they are not the residue of a dark pattern.

05The instrument has already been fogged

This is where my own proposal runs into trouble, and it is worth being blunt about it rather than admiring the idea. Anthropic sells the continuation in bundles, at a discount, and lets you automate the purchase. Both features push dollars away from the moment of truth and toward the moment of planning — which is exactly the moment that carries no information.

Figure 02 The discount ladder pays you to decide early

Pay as you hit the wall $100

$50 bundle $90

$250 bundle $80

$1,000 bundle $70

• A decision made at the limit
• A bet made in advance

The discount is not a trick. It is an ordinary volume discount, and from the company’s side it is obviously correct: prepaid balances smooth demand, reduce checkout friction, and increase the odds that the work continues. But every dollar bought at thirty per cent off, a week early, is a dollar that no longer records anyone’s moment of truth. Anthropic built the measuring instrument and then, for perfectly good commercial reasons, built the thing that fogs the lens.

The same is true of automatic reload, and of the second route out of the wall: a Claude Code user who hits a subscription limit can switch to a Console account and continue on ordinary pay-as-you-go API credits, which land in a completely different ledger. There is even a version where the choice never happens at all — an ANTHROPIC_API_KEY in the environment silently overrides the subscription and bills at API rates without asking.

None of which invalidates the metric. It just means the pure version — first-time, undiscounted, manually confirmed, bought inside the limit notification — is a shrinking subset of a growing number, and anyone reporting it honestly would have to say so.

06The threshold that would settle it

I do not want to invent a magnitude, because no magnitude has ever been published and a made-up one would be worse than none. So the test I would apply is a ratio rather than a level:

Top-up dollars consumed by individual subscribers, as a percentage of individual subscription dollars.

If that ratio sits in the low single digits, the plan is the product and the limit is a nuisance people route around by waiting. If it runs to a meaningful fraction — and especially if it ever exceeds one hundred per cent, meaning people pay more to finish things than they pay to have access at all — then the argument is over. At that point Claude is not a subscription anybody is idly renewing. It is a tool people are buying more of, in the middle of the job, with the free option sitting right there.

Billions of dollars a year of that, from individuals, would be the most convincing evidence of value creation this industry has produced. It would also be the least gameable, because the only way to fake it is to make the limits worse — and making the limits worse is visible, unpopular, and immediately reported. On 29 August 2026, Anthropic announced it would permanently raise standard weekly Claude Code limits by 25% from 14 September. Relative to the temporary 50% boost then in place, that is a 17% reduction, which Anthropic conceded in writing within hours of the announcement. Limits are not a quiet lever.

07The ledger

Three claims, each with the condition under which I am wrong.

08What I left out

Things I wanted and could not confirm, listed so the gaps are visible rather than papered over. No figure exists, at any lab, for revenue from usage top-ups. The current split between API and subscription revenue is unknown; the ~80% business-customer figure is from October 2025 and predates roughly a sevenfold increase in run-rate, so I have used it only as a stale marker. I could not source an exact date for when extra usage became available to Pro subscribers, only that the capability was described for Max subscribers alongside the weekly limits in 2025 and that the discounted bundles page is dated May 2026. And I have deliberately not compared Anthropic’s top-up behavior with OpenAI’s or Google’s credit products, because I could not establish that the three are close enough in mechanics for the comparison to mean anything.

One more, on the reported second-quarter profitability: what Bloomberg described was positive adjusted operating income on preliminary figures, not audited net income, and a widely repeated $559 million profit figure appears to be an internal forecast shared with investors in May rather than a result. I have kept it out of the argument entirely.

09The question itself

What I keep returning to is the wording of the prompt. It does not say your allowance is exhausted, or that your account requires attention. It asks whether you would like to continue. It is the only question in this industry that gets answered with money by the person who actually experienced the value, at the moment they experienced it, with a free alternative in plain view.

Count the yeses. That is the whole audit.