Chinese AI Models Just Hit 46% of US Enterprise Tokens — Here’s Why Devs Are Ditching GPT-5.6

Definition: Chinese-origin models reached a 46% weekly peak of U.S.-organization token volume on OpenRouter, according to a CNBC investigation based on router traffic.
Example: A team can send extraction and summarization to a cheaper open-weight model while reserving GPT-5.6 Sol for ambiguous reasoning or high-stakes coding.
Key takeaway: This is a pricing and routing story before it is a national-model victory lap.
Business impact: Model selection is becoming an inference procurement decision, with quality, cost, hosting, and compliance evaluated together.

The headline sounds like a clean geopolitical upset: Chinese AI models have taken 46% of U.S. enterprise tokens, so developers must be abandoning GPT-5.6. The evidence is more interesting—and more precise—than that framing.

The 46% number comes from a weekly peak in OpenRouter traffic, not a census of every AI request made by U.S. companies. CNBC reported that Chinese-origin models accounted for more than 30% of the tokens routed by U.S. organizations through OpenRouter every week from February 8, 2026 onward, with the share reaching 46% in a later week. The same series put Chinese-model usage at 4.5% in the first half of 2025 and 11% as a trailing twelve-month average. CNBC’s report is the source for the enterprise-token comparison.

That is a dramatic change in routing behavior. It is not proof that 46% of U.S. enterprises, 46% of enterprise revenue, or 46% of sensitive production workloads have switched providers. Tokens are a volume measure. A model can accumulate enormous token share from long-context, repetitive, low-cost tasks without winning the hardest work.

The real story is model routing, not a flag change

OpenRouter sits between developers and model providers. Its traffic therefore captures a decision that benchmarks and vendor surveys often miss: which model an application actually calls when somebody has to pay for the request.

OpenRouter’s own DeepSeek V4 adoption analysis describes a sharp increase in DeepSeek’s share after the V4 release, with agentic workloads driving much of the gain. That matters because agent systems can generate large amounts of context, tool output, retries, and intermediate reasoning. At that volume, a small per-token price difference becomes an architecture constraint.

The practical pattern is not “one company picks China.” It is a workload split:

That is why “developers are ditching GPT-5.6” is directionally catchy but operationally incomplete. A team can reduce GPT-5.6 traffic substantially without removing it. The frontier model becomes the expensive escalation path rather than the default handler for every request.

The price gap makes experimentation rational

The raw economics are hard to ignore. OpenAI’s official GPT-5.6 pricing lists GPT-5.6 Sol at $5 per million input tokens and $30 per million output tokens. The DeepSeek API pricing page lists DeepSeek V4 Flash at $0.14 per million input tokens and $0.28 per million output tokens before cache-hit pricing.

That is roughly a 36× input difference and more than a 100× output difference at the listed rates. Output is especially important for agents because a single user request can trigger many model-generated steps. A cheaper model does not need to win every benchmark to win a budget review; it needs to be good enough on the workload being routed to it.

Z.AI’s GLM-5.2 announcement makes the same case from another angle. The company positions its model for long-horizon coding work, publishes benchmark comparisons, and releases it under an MIT license. Its pricing documentation lists GLM-5.2 at $1.40 per million input tokens and $4.40 per million output tokens—still materially below GPT-5.6 Sol.

The important word is good enough. A price advantage turns into production share only when quality is sufficient for a defined task and the engineering team has evaluations to catch regressions. That favors routing systems, fallback policies, and workload-specific testing over a single universal leaderboard.

Open weights change the deployment question

“Chinese model” can describe at least two different choices:

1. A first-party hosted API operated by a Chinese provider.
2. Open weights deployed through a separate inference provider, hyperscaler, or the company’s own infrastructure.

Those choices can have very different data-governance implications. The model’s training origin does not, by itself, tell you where prompts are processed, how logs are retained, what contract governs the service, or which security controls apply. Conversely, an open-weight model is not automatically safe simply because it can be self-hosted.

A serious procurement review therefore asks four separate questions:

• Capability: Does the model pass the company’s task-level evaluation?
• Economics: What is the cost after retries, tool calls, caching, and long outputs?
• Operations: Can the team monitor latency, failures, drift, and fallback behavior?
• Governance: Where do prompts and outputs run, and what retention and access rules apply?

This is also where the 46% statistic needs restraint. Routing a token to a Chinese-origin open-weight model says something about the selected model and platform. It does not reveal the full security posture of the deployment.

The counter-signal: broad enterprise replacement is not proven

There is a useful check against the most dramatic interpretation. In Ramp’s July AI Index update, model-serving platforms are used as an imperfect proxy for open-source and Chinese-model adoption. Ramp says 5.8% of AI-spending businesses used those platforms in June, up from 4.5% in January. Among those platform users, 85.8% also used OpenAI, 93.2% also used Anthropic, and 96.4% used at least one of the two.

That is not a contradiction. OpenRouter can show Chinese-origin models taking a large share of tokens among a high-volume routing population while Ramp shows that most businesses using model-serving platforms still keep the major U.S. providers. One measures routed token volume; the other measures business-level vendor adoption. They answer different questions.

The more defensible conclusion is that advanced AI users are becoming multi-model buyers. They are willing to use a frontier model where quality is binary, but they are no longer willing to pay frontier prices for every token that moves through an application.

A five-minute migration test for developers

The safe response is not to replace a default model because a market-share headline is exciting. It is to test the workloads that create the bill.

1. Export representative prompts

Take a real sample of production-shaped requests: successful cases, edge cases, long contexts, tool calls, and failures. Remove secrets and personal data before sending anything to an external evaluation service.

2. Define pass/fail criteria before comparing models

Measure factual accuracy, structured-output validity, tool-call correctness, latency, refusal behavior, and cost. For coding agents, include whether the change passes tests—not just whether the model produces plausible code.

3. Route by workload, not nationality

Start with the lowest-risk, highest-volume tasks. Keep GPT-5.6 or another frontier model as a fallback for cases that fail evaluation. The useful policy is “model A for this task under these conditions,” not “country X is always better.”

4. Add an economic guardrail

Track total cost per completed task, including retries and escalation. A model with a low list price can lose its advantage if it requires more attempts or produces longer outputs.

5. Re-run the evaluation after every model update

The 46% story itself is a reminder that model choice is moving quickly. A routing decision that works this month can change when a provider updates pricing, a new open-weight release arrives, or a frontier model becomes cheaper.

What the 46% peak actually signals

It signals that price-sensitive developers have found credible alternatives and are routing meaningful production-like volume to them. It also signals that OpenRouter’s neutral position makes model substitution visible in a way that direct vendor reporting does not.

It does not establish that Chinese models have won every important workload, that U.S. enterprises have broadly abandoned GPT-5.6, or that token share equals revenue share. Ramp’s business-level data argues against a simple mass-exodus story, while OpenRouter’s traffic shows why the pressure is still real.

The strategic threat to OpenAI and other U.S. labs is therefore not just a single Chinese model. It is the normalization of model-level procurement. If developers can separate hard problems from commodity inference, benchmark prestige becomes less valuable than the cost and reliability of completing each task.

That is the change to watch: not whether every team ditches GPT-5.6, but whether defaulting every workload to one expensive model becomes indefensible.

FAQ

Is the 46% figure global market share?

No. It is a weekly peak in Chinese-origin token volume among U.S.-organization traffic measured through OpenRouter. It is not global market share, total U.S. enterprise AI usage, or revenue share.

Does lower price mean the Chinese models are as capable as GPT-5.6 Sol?

Not universally. The case for substitution is workload-specific. Cheaper models can be adequate for extraction, summarization, retrieval-augmented drafting, and routine agent steps while frontier models remain valuable for difficult reasoning, long-horizon coding, and ambiguous tool use.

Should a company send sensitive data to a Chinese-hosted API?

That requires a security, privacy, legal, and procurement review. Model origin, API operator, hosting location, retention policy, contractual protections, and deployment architecture are separate questions.

What should developers do first?

Run a blinded, task-level evaluation on representative prompts, calculate cost per successful task, and add a fallback to a trusted frontier model. Change routing only after the evaluation shows an acceptable quality and governance tradeoff.

The strongest evidence in this story points to a market becoming more selective, not to a single flag replacing another. Token share is the signal; workload-level routing is the mechanism; verification is what keeps the savings from becoming a new source of operational risk.