The end of tokenmaxxing: why AI buyers are routing to cheaper open-weight models

For two years the operating assumption inside every large AI deployment was simple: use more. Staff were encouraged, and sometimes formally rewarded, to run as many tokens through Claude and GPT as they could, on the theory that the productivity upside dwarfed the bill and that the bill would fall anyway. The posture even earned a name — tokenmaxxing. In late June it started to look like a mistake. On 26 June, CNBC reported that a widening set of the largest buyers — Uber, Microsoft, Salesforce and Meta among them — have begun rationing employees' access to advanced models, because the per-token pricing that OpenAI and Anthropic favour had proved, in the words of the reporting, more expensive than it was worth.

The detail that lands hardest is Uber's. The company introduced spending tiers starting at roughly $1,500 a month per employee, with more available on request, after chief technology officer Praveen Neppalli Naga disclosed, via The Information, that it had burned through its entire annual AI budget in four months. That is not a rounding error or a rogue team; it is a disciplined engineering organisation discovering that consumption-priced intelligence, left uncapped, behaves like an open bar. The reflex response is not to stop drinking. It is to find a cheaper supplier.

The bull case for token-metered AI rested on a comforting loop: prices fall relentlessly, so today's extravagant spend is tomorrow's rounding error, so buy now and optimise later. The loop still runs — inference does get cheaper — but the frontier US labs have lately been slower to cut their own prices than the market assumed, and buyers noticed the gap. When the incumbent will not drop its price fast enough, a sophisticated customer does not wait; it re-routes. That is precisely what is happening, and the destination is increasingly open-weight models developed in China.

Flo Crivello, chief executive of the roughly 25-person AI-agent startup Lindy, offers the cleanest illustration. He moved 100% of Lindy's traffic off Anthropic's Claude to DeepSeek — cheaper, open-weight, Chinese — describing the moment the maths turned: "you could see that cost curve go down, like, crash to the ground." He expects to save millions within months and called the switch a matter of survival; one analysis put the inference-cost cut on the migrated routes at around 90%. Crucially, he framed it as reversible — he would switch back to Claude if prices come down, and "until then, we've got options." That last line is the whole story in miniature: loyalty in this market is priced by the token.

The clearest read on the drift is OpenRouter, the API-routing layer many teams use to shop across models. Chinese-developed models now account for roughly 61% of token consumption among its top ten — Qwen from Alibaba, DeepSeek, Kimi from Moonshot, GLM from Zhipu and MiniMax leading the pack. The trajectory is the point:

In eighteen months a rounding error became a majority.

Two forces are compounding. The first is the obvious one — cost, paired with capability that is now good enough for production rather than demos. US startups are building core products, code generation and autonomous agents, on these models, not toy features. The second is subtler and more strategic: open weights. A firm can download the model and run it on its own hardware, with no third-party cloud in the loop. That removes a dependency, a per-token invoice and a supply-chain worry all at once — and it is the second force that turns a procurement decision into an architectural one.

The timing is unforgiving. OpenAI and Anthropic both filed confidentially for historic public offerings in early June, and both go to market on a business model that is now the subject of open dispute. Palantir chief executive Alex Karp called the token-based approach "completely wrong" on 1 July — a pointed line from a company that sells outcomes rather than tokens, but one that lands because the buyers are voting the same way with their budgets.

None of this means the labs are cornered — it means they are responding. Anthropic's Sonnet 5, launched on 30 June at a $2/$10 introductory rate, is engineered to make agent runs cheap enough to keep them in-house; OpenAI's lower-cost Terra tier — previewed 26 June, though still gated to approved partners — is aimed at the same pressure. The efficiency turn is not only reshaping how enterprises buy — it is reshaping what the leading labs choose to ship.

The honest read is not that the frontier labs have lost. It is that the pricing power everyone assumed they held — the ability to meter intelligence by the token and trust that usage would only rise — has become negotiable. Buyers have discovered a credible alternative, and a credible alternative is all it takes to turn a seller's market back into a market. For companies about to ask public investors to underwrite a token-metered future, that is the number that matters.