Something unusual is happening in the AI industry: the price of the product is falling while the cost of building it climbs. Both trends accelerated this week.

On 13 August, Google launched Gemini 3.7 Flash, aimed at coding and automated business workflows, at an introductory $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026 — half the launch price of its predecessor. The same day, OpenAI previewed Ultrafast, an API tier running its flagship model up to fourteen times faster than standard processing, generating up to 750 output tokens per second, using hardware from Cerebras.

Meanwhile DeepSeek, the Chinese developer whose low prices unsettled the market last year, has been raising its own.

What a token price actually is

For readers who do not buy this stuff: a token is roughly three-quarters of a word. Providers charge separately for text sent in and text generated out, with output costing several times more because it requires far more computation.

At $0.75 per million input tokens, processing an average novel costs well under a dollar. That is the number that matters for anyone deciding whether to build a product on this: it determines whether a service that calls a model thousands of times a day is viable or ruinous.

Speed as the new battleground

The Ultrafast announcement points at where competition is moving. For a chatbot, generating text faster than a person reads is enough. For an agent doing a hundred sequential steps, or a voice system that must reply without an awkward pause, latency is the constraint that decides whether the product works at all.

That fourteen-times figure comes from specialised chips rather than a better model — the same intelligence, delivered faster. It is a reminder that the industry’s competitive edge increasingly sits in silicon and infrastructure rather than in the models themselves.

The arithmetic problem

Here is the tension. Prices are falling. At the same time, the large technology companies’ AI purchase commitments are reported to be approaching $1.5 trillion, chip manufacturers are running near capacity, and Anthropic is reported to be projecting roughly $190–200 billion in 2028 revenue as it prepares an IPO — against a $47 billion run rate disclosed in May.

Those two things can only be reconciled by volume: prices per unit fall, usage rises faster, revenue grows anyway. That is the standard software story, and it has worked before. It requires demand to keep expanding at a rate that justifies infrastructure spending committed years ahead of the revenue.

Introductory pricing is also worth reading carefully. Half price “through the end of 2026” is a customer-acquisition offer with an expiry date attached, and the industry has form in this: subsidise adoption, establish dependency, revisit the price later. DeepSeek raising its prices is the same story from the other end.

Why this matters outside the industry

Two reasons. Cost determines where these systems get deployed — falling prices are why they are appearing in customer service, document processing and public-sector systems that could not have justified them two years ago, with consequences for employment that arrive quietly.

And the capital involved is now large enough to matter to everyone’s pension fund. When a handful of firms commit sums approaching the annual output of a mid-sized economy on the expectation of demand that has not yet materialised, the question of whether that expectation is sound stops being a technology question.

Sources

Google and OpenAI announcements of 13 August 2026 as reported by industry trackers; Reuters reporting on Anthropic’s revenue projections, 14 August 2026; contemporaneous coverage of AI infrastructure commitments and SMIC capacity.