What happened
Sonnet 5 arrived on 30 June at an introductory price: two dollars per million input tokens and ten per million output. The launch announcement carried an end date, 31 August, and the standard rate meant to take over from 1 September: three dollars in, fifteen out. The deadline was stated plainly, so it went into second-half forecasts. Including ours.
On 10 August Anthropic added one sentence to that same announcement. The introductory price becomes the standard price, and the September increase will not take effect. The pricing documentation now says the same thing, with no deadline and no asterisk.
Two weeks later a second fact arrived, this time from the demand side. On 23 August the Financial Times reported data from Ramp, the company that handles corporate cards and bill payments for more than seventy thousand businesses. It showed that Fable 5, Anthropic's most capable and most expensive model, accounted for roughly 6% of the tokens bought from that vendor one month after launch, and 11.4% of spending on its models. The older Opus 4.8 took 28% over the same period. In its 12 August edition Ramp calls this a crack in the AI thesis: the best model on the market did not collect the money it was expected to collect.
The price list suggests why. A million output tokens from that one vendor costs you fifty dollars or ten, depending solely on which model name you put in your configuration. Input runs to ten dollars or two. Everything else — the contract, the endpoint, the integration — stays the same.
Our read
Start with a correction, because it is ours. In July we wrote about the axes along which model price lists had come apart, and among them we listed the September Sonnet increase as a rise already in the calendar, worth putting into your forecast now. That line has been void since 10 August. If somebody reserved budget for it, they can release it, and that matters less than the reason this has to be written today.
A date in a launch announcement looked like a fact about the price list. It was a statement of intent. The difference did not matter for as long as every such statement was honoured.
We flag this explicitly as our interpretation, not an established finding: Anthropic gave no reason for withdrawing the increase and does not publicly connect it to the Ramp data. Putting the two events side by side is our reading. The reading goes like this: the increase was lifted from the tier that carries the volume, in the same month that the tier above it failed to defend its price. That does not look like a gesture towards customers. It looks like a test the vendor ran and lost.
It is worth seeing what actually broke. Not budgets: companies are spending more on AI, not less. What broke is the assumption that a better model sells itself, and that you can therefore charge any multiple of the price of the model next to it. The market replied that the premium has a ceiling, and set it lower than the price list assumed.
The reason is more mundane than thrift. Almost nobody can show what the more expensive model actually adds on their own work. Without your own test on your own tasks, the only comparable number on the table is the rate per million tokens, so you optimize the one number you have, because there is no second one. A public benchmark ranking is not that second number, because it does not answer whether your service tickets will be classified better.
Why this matters
Private Equity
An investment thesis resting on AI spend per employee rising in step with model quality now has a counter-example, and it is worth reading before the next committee. From the same Ramp edition: the median firm spends $11.95 per employee on AI, the top ten percent $650, and the top one percent $7,400. This is not a market growing evenly. It is a narrow group spending heavily and a long tail spending close to nothing.
The practical question for a portfolio-company review is single and cheap: what share of the token bill sits on the most expensive model tier, and which business decision rests on it. "We use the best model" is not a cost plan, it is a default nobody changed after the pilot.
Enterprise
The good news is short: a budget line that was due to grow from September will not grow, so the working tier can be planned into the next cycle without a reserve against that increase.
The longer conclusion is about what this says about price-list control in general. The price of the model moved twice this year in opposite directions, each time on a decision you had no part in, on a date you did not set. The consequence is the same for a rise and for its cancellation: the model identifier has to be a line in your configuration that somebody changes deliberately, not a default value inside a client library. Alongside it, one set of your own test tasks, run at every change of model tier. Without it you are negotiating the price of something whose quality you cannot measure in your own context.
SMB / mid-market
Here the outcome is simply favourable and calls for no defensive move. The model tier most of the real work in a mid-sized company actually runs on became cheaper permanently, rather than until the end of August. If anyone was planning a migration to a cheaper provider before September purely because of that increase, the premise is gone.
What is worth taking from this for the longer term is something else. The Ramp data shows that companies paying for AI keep the most expensive model for a narrow set of tasks and hand the rest to cheaper ones. That is not cutting corners on quality, it is ordinary tool selection — and it is exactly how we pick the model layer in deployments for mid-sized companies: first the list of tasks and which of them genuinely need the strongest model, then the price list.
One move this week
Take the last invoice from your model provider and break it down by model identifier rather than by project. For the most expensive line, answer one question: which specific task sits on it, and what breaks if it drops a tier. If you do not know, run that task on the cheaper model against twenty real cases from last week and compare the results by hand. An hour of work gives you a number you do not have today, and that number, not the price list, should decide what you pay extra for. Describe your case: mailto:[email protected]?subject=Rozmowa%20z%20Aurora%20AI