The model mix is now mandatory.

Prices checked July 15, 2026. One caveat on that GLM row, and it generalizes. Open weights do not have a price, they have a market. Around twenty hosts serve GLM-5.2, and on a single day in early July their listed input rates ran from $0.93 to $3.00, with output reaching $10.25 on premium latency tiers. The cheapest route also carried the tightest output ceiling, 32,768 tokens, which silently truncates a long agent run. The sticker is not the bill. Check both against your own workload before you build a business case on either, and check them again in a month, because that GLM input floor has dropped by a third in ninety days.

When the spread between tiers runs 50x, routing becomes the biggest cost lever most teams have. Bigger than caching. Bigger than prompt compression, and a lot less fiddly than either.

You can do this arithmetic yourself, which is better than trusting anyone's savings percentage. Take a million input tokens at Opus 4.8's $5.00 and the bill is $5.00. Move 70% of them to GLM-5.2 at $0.93 and you pay $0.65 plus $1.50, so $2.15. That is a 57% cut from one routing decision, before you touch caching. Push the cheap share to 80% and it approaches two thirds.

Routing is not free, and any honest version of this argument says so up front. You are adding a component to the request path: a classifier or rule layer that costs milliseconds, occasionally misroutes, and becomes one more thing that can fail at 3 a.m. Against a 50x price spread that trade is usually worth making. It is still a trade.

The uncomfortable part is how few teams make it.

Put those two findings side by side. The industry knows the answer and has not acted on it. Which is not really a technology problem. Nobody gets promoted for switching a working pipeline to a cheaper model, and everybody gets a bad quarter when the switch goes wrong. The gap is organizational, and it is where the money is.

The analysts see the same thing coming.

The cheap tier did not appear from nowhere. It was published, deliberately, under permissive licenses.

Calling China a token exporter needs unpacking, because it is not exporting tokens the way it exports steel. It publishes weights. Anyone can download them, and the inference happens on whatever hardware the buyer chooses, in whatever country. What crosses the border is the capability, once, for free. The revenue shows up later and somewhere else: cloud contracts, enterprise deployment, paid endpoints for teams who would rather not run their own.

DeepSeek alone doubled from 9% to 18% of platform token share between January and June. The tipping week was February 9 to 15, when Chinese systems processed more tokens than American ones for the first time. What makes that worth your attention is not the flag. It is that developers switched fast, in production, on price, which tells you the switching cost is lower than most procurement conversations assume.

The strategy behind it is not charity. Alibaba runs Qwen as top-of-funnel for Alibaba Cloud: open weights seed adoption, paid inference captures the tail. Qwen passed a billion cumulative downloads on Hugging Face, overtaking Llama. Zhipu, now listed in Hong Kong, ships GLM under MIT and sells enterprise deployment.

Export is still the right word for it. A model published under MIT and served by any host anywhere is an inference product that crossed a border without a customs declaration, and without leaving anyone in Washington a switch to flip.

If you have read this far you have probably seen a chart of this shift with a red line going up. Here is what those charts leave out, and it changes the conclusion.

There are two markets here, not one. A commodity lane, where the cheapest capable model takes the bulk work: classification, extraction, retrieval, the jobs measured in billions of tokens a month. And a premium lane, where reliability and genuinely hard reasoning still command a markup that buyers keep paying. Headlines about collapsing American share are measuring the first lane while implying something about the second.

Your workload lives in one of them. Realistically, in both.

A cheap token price says nothing about data governance. And with Chinese models, one decision dominates every other: how you run it.

Self-hosted open weights send nothing back to the developer. Calling a China-hosted API can place your data under Chinese law, since Chinese companies are subject to the National Intelligence Law and can be compelled to assist state intelligence work.

That splits into three practical paths:

The open license is what makes path one possible, and path one is the reason open weights matter more than a cheap hosted endpoint.

This section is not a warning about China. It is a warning about single points of failure, and it happens to apply symmetrically.

Policy discussions with named labs, not a signed decree, and no comment from the ministries or the companies when Reuters asked. Treat it as weather, not climate.

The symmetry is still the point. In June, Washington export-controlled two Anthropic models and put OpenAI's GPT-5.6 Sol behind customer-by-customer approval. Both capitals now treat frontier models as strategic assets, and both have shown they will act on that view with short notice.

Which cuts a specific way for planning. Weights already published cannot be recalled: DeepSeek V4 and GLM-5.2 are on disk, worldwide, permanently. Future ones can be withheld by either government. So build on the tier you have already downloaded and treat next year's releases as a bonus rather than a dependency. That is a different posture than betting your roadmap on a pricing trend continuing.

1. 1

   Measure before you migrate

   Start with a measurement, not a migration. Pull last month's inference bill, segment it by task type, and look at what the money is actually buying. Most teams find the bulk of their calls are classification, extraction, or retrieval, all of it running on frontier models for no better reason than that is what got wired in first.

2. 2

   Evaluate on your own traffic

   Then run your own evaluation, and run it on your traffic rather than on a leaderboard. The version worth copying is deliberately unsophisticated. Pull 200 real requests from last month's logs, weighted toward whatever task type dominates your bill. Run each one through the incumbent model and the candidate. Have someone who knows the domain grade both outputs blind, on a scale that means something to your business, not on a scale of one to five. Count the disagreements. If the cheap model loses on fewer than a handful and none of those are expensive to get wrong, you have your answer, and you have it from your own traffic instead of a benchmark someone else ran.

   Published benchmarks tell you a model is capable in general. GLM-5.2 leading the Artificial Analysis Intelligence Index among open weights is a fact about GLM-5.2. It is not a fact about your pipeline.

3. 3

   Route the easy majority down

   Route the easy majority down and keep the hard minority where it is. Measure quality on both sides, because silent quality regression is the real risk in production routing and it does not announce itself in the bill. It shows up in retries, in support tickets, in a reviewer quietly rewriting output nobody logged. This is the Baseline and Bearing part of our AZIMUTH method doing ordinary work: you cannot route on evidence you never collected.

   What does the exercise itself cost? For a team with one production pipeline, the realistic answer is a week: a day to segment the bill, two to build and grade the eval set, two to wire a route and watch it. The expensive part is not the engineering, it is finding someone with the domain knowledge to grade outputs without flinching, and that person is usually busy. Budget for their time, not for tooling.

4. ■

   Write the kill criterion first

   Write the kill criterion before you start, not after the first bad week. Ours for this kind of work reads roughly like this: if the cheap tier costs more in retries, review time, and escalation than it saves in tokens, we stop, we say why in writing, and the incumbent model keeps the traffic. Deciding that in advance is what stops a routing project from turning into a sunk-cost argument three months in, when somebody has a dashboard to defend.