The dial goes to max
How much a model thinks used to be decided at training time and sold as a product tier. It is becoming a request parameter, and that moves the spending decision to whoever knows which questions are hard.
V4-Pro reached general availability on 16 August with a reasoning-effort setting at low, high or max, agent upgrades, and native support for a rival's API.
Tiers were scaffolding
A model family with fast, standard and deep variants is a taxonomy imposed on a continuum. It exists because inference cost had to be exposed somehow, and a SKU was the only handle available.
A dial is the same idea with the granularity moved. Instead of choosing a model and living with its economics for every request, you choose per request — which means the cost decision now sits with the caller, at the moment they have the most information about what the query is worth.
It rewards teams that know which of their queries are hard, and it punishes the ones that do not.
The same collapse is happening everywhere this week
Off-peak tokens now cost half of peak, with the windows published — pricing moved from a rate card to a schedule. The Imagen 4 endpoints closed today in favour of the general model — modality moved from an endpoint to a parameter.
Three tiers into one dial, one price into two, three image endpoints into none. The pattern is the same in each case: what used to be a product distinction is becoming an argument you pass in.
What that does to comparison
It makes benchmark tables less meaningful and TCO harder to state. A model with a reasoning dial does not have a latency or a price, and the number a vendor quotes is now a choice about which setting to quote.
The buyers who come out ahead are the ones instrumenting their own traffic. If you know that four per cent of your queries need max and the rest need low, you can price your product. If you do not, you will discover your bill is a function of something you never measured.
Digital Applied — AI Model Releases: August 2026 Tracker and Dated Ledger → · LLM Stats — AI Updates Today (August 2026) →