China’s models grew up. Your AI bill should notice.

Two releases define the year so far. In February, Zhipu (Z.ai) shipped GLM-5 — a 744-billion-parameter open-weight model, trained entirely on Chinese silicon, that benchmarks alongside Anthropic’s frontier coding models. In April, DeepSeek previewed V4: a 1.6-trillion-parameter system with a 1-million-token context window as standard, weights on Hugging Face under MIT, and the official release landing this month. Both are serious production models, not curiosities. And both are reshaping what a token should cost.

The twist: cheap is putting its prices up

The story used to be simple — Chinese labs racing each other to zero. That race is over. Zhipu has raised prices twice in five months and its stock jumped on both announcements; investors read the hikes as proof the demand is real. DeepSeek’s peak-hour doubling is the same signal in a different shape: when a lab starts charging by the clock, inference demand — not training — has become its binding constraint. The era of flat, infinitely scalable, loss-leader tokens is closing.

And yet the gap barely moved. Even after two increases, GLM-5.2 lists at roughly a sixth of comparable US output pricing; DeepSeek V4 Pro, even at its doubled peak rate, stays far below any Western frontier list price. A Deutsche Bank analysis this June put V4 at roughly 1.5% of the cost of Anthropic’s flagship for the large majority of everyday tasks. Cheap got slightly less cheap. It did not stop being cheap.

Where the gap actually lives

Price is only half the ledger; the other half is what a model gets right first time. On SWE-bench Verified — the closest thing coding has to a common exam — DeepSeek’s best V4 configuration resolves 80.6% of tasks, against roughly 88–95% for the Western frontier. Eight to fifteen points doesn’t sound like much until you translate it: on a hundred real tickets, that’s eight to fifteen extra failures to catch, re-prompt or hand to a human. The dollar arithmetic still favours the cheap model — even three retries of V4 cost pennies against one frontier pass — but retries are paid in waiting, and a senior engineer’s afternoon is the most expensive token there is. That’s why the winning pattern in every stack we’ve audited this year is a mix, not a swap.

A day in the life of a surge-priced token

DeepSeek’s peak windows follow the Beijing working day — which is quietly good news for everyone west of it. For a London team, the 2× hours land between one and four in the morning and before mid-morning; for the US East Coast they sit overnight. Ordinary Western business-hours traffic mostly rides the cheap side of the clock, and scheduled work — evals, batch classification, data generation — can be pinned there deliberately. Timezone arbitrage is now a line in the runbook.

Six months, four price moves

Put the year’s announcements on one line and the direction is unmistakable: the two labs that taught the world to expect ever-cheaper tokens both spent 2026 repricing upwards — in opposite styles. Zhipu raised the sticker; DeepSeek kept the sticker and charged for the busy hours.

Whether any of this belongs in your stack is a measurement question — the kind a two-week Reality Check answers with your workloads, your data rules and your numbers, not a benchmark chart.