For a growing share of real company work, teams have stopped treating the frontier lab as the only serious choice. They ask a smaller, more precise question first. Does this specific task actually need the most expensive model, or is a cheaper one good enough?
That question is quietly reshaping how AI bills get paid.
CNBC reported on July 7 that Chinese-origin AI models now account for more than 30% of U.S.-origin token usage routed through the platform OpenRouter, and have every week since February 8, 2026, peaking as high as 46%. The prior 12-month average was around 11%. In the first half of 2025 it was roughly 4.5%. That is a fast climb in a short window. Price appears to be the strongest immediate driver, alongside cheaper models that now hold up well enough on a lot of the work.
Why price is doing the work
Look at one concrete pair. OpenAI's GPT-5.6 Sol, its current flagship, costs $5.00 per million input tokens and $30.00 per million output tokens on OpenAI's own pricing page. DeepSeek-V4-Flash lists input at $0.14 per million tokens on a standard cache-miss request. On that specific pair, the cheaper model runs roughly 36 times less per input token.
The peak weekly share of U.S.-origin token usage running on Chinese-origin AI models through OpenRouter, up from a prior 12-month average around 11%.
Source: CNBC analysis of OpenRouter data, 2026Now, that 36x figure is one extreme example, not a blanket rule. Not every Chinese model is 36 times cheaper. A broader read comes from Justin Summerville at OpenRouter, quoted by CNBC: "Open-source Chinese models are consistently 60% to 90% cheaper than the leading offerings from Anthropic and OpenAI." Sixty to ninety percent is the range that shows up across models. The 36x pair sits at the far edge of it.
Harpreet Arora, who leads agentic infrastructure at Vercel, put the behavior plainly to CNBC: "Price is doing the work here. When a task doesn't need the best model, teams are beginning to route it to the cheapest one that's good enough, and the recent wave of models coming out of China is winning that trade."
Read that again. The cheapest model that is good enough.
That is a portfolio decision, and it is one worth your attention early.
This is not a niche experiment anymore
The cheap models are not just cheap. Some are competitive at the frontier.
Z.AI released GLM-5.2 on June 13, 2026, to its Coding Plan, with open weights and standalone API access following on June 16. It is a long-horizon coding model with a one million-token context window. Standard GLM-5.2 is priced at $0.95 per million input tokens and $3.00 per million output on Vercel's AI Gateway. Set that against GPT-5.6 Sol at $5.00 and $30.00. You are looking at an open-weight model built to compete with frontier systems on coding and long-horizon agent work, at a fraction of the token price.
Adoption moved fast. CNBC reported, citing Vercel, that GLM-5.2 saw the fastest uptake of any model Vercel tracked in 2026. In its first full week after launch, daily token volume grew about 27 times and customer count about 80 times.
The platform data underneath the trend backs the direction. OpenRouter's own year-in-review shows DeepSeek was the single largest model author by token volume between November 2024 and November 2025, at 14.37 trillion tokens, ahead of Alibaba's Qwen at 5.59 trillion and Meta's LLaMA at 3.96 trillion. OpenRouter noted that Chinese open-source models "reached nearly 30% of total usage among all models in some weeks."
A snapshot of provider-level weekly volume that CNBC reported: DeepSeek at 17.6%, Qwen at 13.9%, and Anthropic at 14.8% of U.S.-origin models. Those percentages move week to week, so treat them as a photograph, not a fixed score. The story they tell holds either way. Cheaper models are absorbing serious volume.
What this actually means for the leader who feels behind
Here is the change in thinking. The advice you keep hearing is to just use ChatGPT, or just use Claude. That advice treats the model choice as a brand loyalty question. The teams cutting the model cost of their routine work are treating it as a routing question instead.
They ask, task by task: what does this job require?
- Drafting a first-pass internal email, summarizing a meeting transcript, tagging support tickets. These rarely need the most expensive model. A good-enough cheaper model handles them.
- Reasoning through a legal edge case, writing production code your business depends on, a customer-facing answer where a wrong response costs you. These can justify a more capable model. But paying more does not make the answer safe on its own; high-stakes work still needs testing, guardrails, and a qualified human review.
You do not have to switch anything today to benefit from understanding this. The skill being rewarded is judgment. Knowing which task goes to which model, and why. That is a per-task question now, and the people who can answer it are quietly spending far less for the same output on routine work.
The skill being rewarded is judgment. Knowing which task goes to which model, and why.
Where the real judgment lives: it is not only about price
Cheaper open-weight models carry different considerations than a frontier lab's hosted service. Data handling, where your inputs go, model provenance, and country of origin all become live questions. An open-weight model is often run by a third-party host, so your inputs still leave your systems unless you have the right data agreement in place, or you run the model inside your own environment. A cheaper token is not automatically the right token for sensitive data.
So the muscle this whole change rewards is discernment. It is neither tool loyalty nor blind cost-chasing. It is choosing the right model for the task, weighing capability against price against data risk, and being able to explain the choice.
In The 7 Levels of AI Proficiency, that is the jump from someone who just picks the default tool to a Level 3 Critical Thinker: the operator who trusts but verifies, who asks what a task actually needs before spending on it, and who can defend the decision to a board. The token-routing story is a live example of that level in action. The people saving money are not the ones with the fanciest tool. They are the ones with the clearest read on which job needs what.
A next step
Pick one AI task your team runs every week. Ask three questions about it. Does this job actually need the most capable model? What would break if it ran on a cheaper one? And if this task touches sensitive data, where would that data go? You do not have to change a tool to answer those. Answering them well is the proficiency that separates the teams spending less on routine work from the ones paying frontier prices for it.
Related reading: Level 3: The Lieutenant (Critical Thinker).
Sources
- Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge (CNBC)
- DeepSeek API pricing
- OpenAI API pricing
- GLM-5.2 model page (Vercel AI Gateway)
- OpenRouter State of AI
Frequently Asked Questions
Does this mean I should switch my company to a Chinese AI model?
No. The lesson is not "switch." It is that model choice is becoming a per-task portfolio decision rather than a single brand pick. Routine, low-stakes work can often run on a cheaper model. High-stakes work still warrants the premium one. And any model handling sensitive data needs a data-handling review first, regardless of country of origin.
Are these cost figures reliable?
The model pricing is drawn from first-party pages: OpenAI's pricing page, DeepSeek's pricing docs, and Vercel's model listing. The U.S.-origin token-share percentages (the 30% to 46% figures) come from CNBC's July 7 analysis of OpenRouter data and are CNBC's own cut of that data, not a figure OpenRouter publishes directly. Treat the pricing as verified and the share percentages as a reported analysis.
Why are the cheaper models so much less expensive?
The reporting does not settle the full "why," so this is worth holding loosely. What the data shows is the effect: several Chinese open-source models are priced 60% to 90% below the leading Anthropic and OpenAI offerings, and teams are routing good-enough tasks to them to save money.
What is the one thing to do with this?
Look at your own AI usage and separate it into two buckets: routine tasks and high-stakes tasks. That split is the start of thinking about model choice the way the teams saving money already do.
Find your AI Proficiency level
The free 7 Levels assessment places you across seven stages of AI capability. Under ten minutes. Research-backed scoring.