The end of 'tokenmaxxing': enterprises ration AI spend

Companies including Uber, Microsoft, Salesforce and Meta have moved to ration employees' use of advanced AI, CNBC reported on 26 June. The shift has a name: the end of 'tokenmaxxing' — the practice of rewarding staff for burning as many tokens as possible, with no return-on-investment guardrails. The verdict from buyers is blunt: the token-payment model favoured by Anthropic and OpenAI proved more expensive than it's worth.

Uber is the cautionary tale. CTO Praveen Neppalli Naga said the company blew its entire annual AI budget in four months, then imposed spending tiers starting at $1,500 a month per employee, more on request. Startups are going further. AI-agent firm Lindy shifted 100% of its traffic off Anthropic's Claude to China's DeepSeek; CEO Flo Crivello expects to save millions within months, and says he would switch back if US prices fall.

The money is visibly moving. On OpenRouter, a major API-routing platform, Chinese-developed models now account for roughly 61% of token consumption among the top 10 — a near-total reversal in eighteen months.

Leaders include Qwen (Alibaba), DeepSeek, Kimi (Moonshot), GLM (Zhipu) and MiniMax. Firms switch for cost and capability — and because open weights let them download a model and run it on their own servers, with no third-party cloud. US startups are now building core products, from code generation to autonomous agents, on these models rather than just demos.

The timing is awkward. OpenAI and Anthropic both filed confidentially for IPOs in early June, and their token-metered business model is now under open scrutiny — Palantir CEO Alex Karp called the token-based approach 'completely wrong' on 1 July. Both labs are already cutting agent costs — Anthropic's Sonnet 5 at an introductory $2/$10, OpenAI's cheaper Terra tier — a direct answer to buyers routing around them.