Tokenmaxxing is over: Chinese open models are now 61% of top API traffic

For two years the trillion-dollar labs sold enterprises one story: spend. Reward staff for burning as many tokens as possible, skip the ROI guardrails, because token prices always fall. On 26 June 2026 that story broke in public. Buyers stopped tokenmaxxing, started rationing, and the money began leaking to cheaper Chinese open-weight models. The spend-at-all-costs era that built OpenAI's and Anthropic's ~$1T valuations is cracking — and it's cracking at the worst possible moment for both.

Ask Uber. CTO Praveen Neppalli Naga watched the company blow its entire annual AI budget in four months, then slapped on spending tiers starting at $1,500 a month per employee — more only on request. Uber isn't the outlier. Microsoft, Salesforce and Meta have all moved to ration staff use of advanced AI, because the token-payment model Anthropic and OpenAI favour turned out, in CNBC's words, more expensive than it's worth. Tokenmaxxing — paying people to consume AI with no ceiling — lasted exactly as long as the free-money mood did.

The rationing is now a pattern, not a panic:

The sharpest data point isn't a big-cap rationing memo — it's a 25-person startup voting with its API keys. Lindy, an AI-agent company, moved 100% of its traffic off Anthropic's Claude to DeepSeek: cheaper, open-weight, Chinese. CEO Flo Crivello didn't hedge. "You could see that cost curve go down, like, crash to the ground," he said, expecting to save millions within months and flatly calling the switch "a matter of survival." One analysis put the inference-cost cut on the migrated routes at ~90%. Crucial nuance the doom-takes skip: he'd switch back to Claude if prices come down. "Until then," he said, "we've got options." That's not ideology. That's a spreadsheet.

And it's not one contrarian founder. On OpenRouter, the big API-routing platform, Chinese-developed models are now ~61% of token consumption among the top 10 — up from under 1.2% in late 2024 and ~51% as recently as April. Qwen, DeepSeek, Kimi, GLM and MiniMax are the leaders. US startups aren't slotting these into throwaway demos, either; they're building core products — code generation, autonomous agents — on them. The open weights are half the draw: download the model, run it on your own servers, no third-party cloud, no metered tap someone else controls.

Here's the punchline. Both OpenAI and Anthropic filed confidentially for historic IPOs in early June — and the business model they're marketing to public investors is the exact token-metered meter buyers are now routing around. You don't have to be a bear to see the tension. Palantir CEO Alex Karp isn't subtle about it: on 1 July he called the token-based model "completely wrong." When a fellow enterprise-software CEO is torching your pricing model the same month you're courting the public markets on it, that's not noise — that's the debate the roadshow will have to answer.

Now the fair part, because the doom take overshoots. This is a standoff, not a defeat. At the very top of the capability curve, US frontier models are still the best you can buy, and Crivello himself says he'd come back the moment the price is right. The labs also aren't sitting still: Anthropic's Sonnet 5 at an intro $2/$10 and OpenAI's cheaper (still-previewed) Terra tier are direct, obvious answers to precisely this pressure. Cut the price of running agents and a chunk of that leaked traffic walks straight back through the door.

But that's the whole point, isn't it. "Prices always fall" was supposed to be automatic — the reason you could tokenmax without a care. Instead the leading US labs got slow to cut, buyers built the muscle to route around them, and open-weight rivals took the majority of top-10 traffic before anyone in a corner office called it. The efficiency turn isn't a blip in procurement; it's reshaping product strategy at the top of the market. Tokenmaxxing is over. What replaces it is a market where the meter has to justify itself — and that's a far harder story to sell to an IPO.