The 46% Problem: Chinese Open Models Now Route More Tokens Than America's

What happened

On July 7, CNBC published an investigation into where American developers are actually sending their inference traffic, and the answer is not the place most of the industry's capital has been pointed. Using data from OpenRouter — the model gateway that sits between applications and dozens of providers — CNBC reported that Chinese-origin AI models have accounted for at least 30% of US-originating token volume every single week since February 8, 2026, with a weekly peak near 46%.

The trajectory is the story. Per the reporting, that share averaged roughly 11% over the prior twelve months, and just 4.5% in the first half of 2025. Whatever this is, it happened in about eighteen months, and most of it happened in the last six.

Forkast, summarizing the same dataset, put the vendor breakdown at DeepSeek 17.6% of routed tokens and Alibaba's Qwen 13.9% — with Chinese-origin models collectively at 46.4% against 35.7% for US-origin models, and Anthropic, the largest single American provider on the platform, at 14.8%. Those figures describe a specific week rather than a running average, and they are the most aggressive framing available; the durable claim is the floor, not the peak.

I could not fetch CNBC's article directly — the site returns HTTP 403 to automated requests — so everything above comes from three outlets independently quoting its numbers. Where they disagree, I say so below.

The number behind the number

The mechanism is not mysterious. Justin Summerville of OpenRouter's data and analytics team told CNBC that open-source Chinese models run 60% to 90% cheaper than leading Anthropic and OpenAI offerings, and the published rate cards make that look conservative.

MLQ lists GPT-5.5 at $5 per million input tokens and $30 output, and Claude Opus 4.8 at $5/$25. Against that: Zhipu's GLM 5.2 at $1.40/$4.40, MiniMax M2.5 at $0.30/$1.10, and DeepSeek V4 Flash at the bottom of the market. Here the sources split — MLQ quotes DeepSeek V4 Flash at $0.09/$0.18, while Forkast reports $0.14 input. Either way the ratio against a $5 input price is somewhere between 35x and 55x, and MLQ's "55x cheaper" characterization is directionally sound even if the exact denominator isn't settled.

The second half of the mechanism is that the cheap models stopped being bad. MLQ reports GLM 5.2 scoring 74.4 on the FrontierSWE benchmark against Opus 4.8's 75.1, and 76.8 on MCP-Atlas against Opus 4.8's 77.8. Those are single-benchmark, vendor-adjacent numbers and deserve the usual skepticism. But a sub-one-point gap at a fifth of the price is not a rounding error in a procurement conversation — it is the whole conversation. And GLM 5.2 ships open-weights, which means a company with its own GPUs pays no per-token price at all.

Why the workload mix matters more than the headline

Buried in MLQ's account is the variable that actually explains the shift: programming workloads grew from roughly 11% of OpenRouter usage in early 2025 to more than 50% by mid-2026.

Coding agents are the most token-hungry workload in production software. An agentic loop that reads a repository, proposes a patch, runs tests, and iterates will burn millions of tokens on a task a human would describe in one sentence. That is precisely the regime where a 20x price difference stops being a line item and becomes an architectural decision. It is also the regime where "close enough on the benchmark" is easiest to tolerate, because the agent's own test suite catches the failures.

So the token-share number is partly a story about Chinese models winning, and partly a story about which workload grew. The cheap models didn't take a fixed pie — they took the fastest-growing slice of a pie that changed shape underneath everyone.

The self-inflicted wound

Timing did the rest. MLQ notes that Anthropic's Fable 5 and Mythos 5 were suspended on June 12 under export-control pressure and restored July 1, and that OpenAI limited a model rollout at US government request at the end of June. AI Weekly, citing Vercel, reports that GLM 5.2 posted the fastest adoption of any model on that platform in 2026 — roughly 27x daily token growth and about 80x customer growth in its first full week. (The "27x/80x" figures describe one vendor's platform in one week off a near-zero base, which is how most 80x numbers happen. Treat them as a signal of direction, not magnitude.)

Read those two facts together and the picture is uncomfortable. American frontier labs spent three weeks in mid-2026 being intermittently unavailable or rollout-constrained for policy reasons. Chinese open-weights labs, by construction, have no such chokepoint: once weights are published, no agency can un-publish them. Availability is a feature, and for a stretch this summer the US models didn't have it.

Palantir CEO Alex Karp called the US labs' per-token business model "broken." He has obvious commercial reasons to say so. That does not make him wrong about the direction of gross margins.

What this number does not say

Four things, and they matter enough that the headline should be read as narrower than it sounds.

OpenRouter is not the enterprise. It is a gateway favored by developers, indie builders, and cost-sensitive startups — exactly the population most likely to chase a 20x price cut. Large enterprises overwhelmingly buy direct, through Bedrock, through Vertex, or through Azure. None of that traffic appears here. Calling this "US enterprise token usage," as several outlets have, overstates what the pipe measures.

Tokens are not revenue. A 46% token share at a fifth of the price is roughly a 12% revenue share. DeepSeek routing more tokens than Anthropic tells you very little about which company is collecting more money — and the CNBC-adjacent Ramp figures, which AI Weekly renders as Anthropic $4,811, OpenAI $3,357, Zhipu GLM $544 in business spending, point the other way. The unit and period on those dollar figures are not clear from the secondary sources, so I would not lean on them, but their ordering is the point: on spend, the American labs are not close to losing.

Open weights invert the metric. Every company that downloads GLM 5.2 and self-hosts disappears from OpenRouter entirely. The gateway share undercounts Chinese model adoption, possibly badly.

Data governance is unresolved. Routing tokens through DeepSeek's hosted API and running DeepSeek's weights on your own hardware are different acts with different risk profiles, and the coverage tends to blur them. The self-host path is the one that makes export controls irrelevant, and it is also the one that makes the security objection mostly go away.

Hype vs. real

Real: the price collapse, the benchmark convergence, the coding-workload shift, and the June availability gap. Those are four independent forces pushing the same direction, and none of them are reversible by press release.

Overstated: that this is a leaderboard flip. It is a gateway flip, on a platform selected for price sensitivity, measured in the one unit — tokens — that flatters the cheap option by construction. Anthropic and OpenAI are still, per every spend proxy available, capturing most of the money.

Genuinely unresolved: whether frontier labs can defend a 20x price premium on capability alone once open weights are within a point on the benchmarks that buyers check. Anthropic's answer so far has been to cut Sonnet 5's price and lean on agentic reliability. Whether reliability is worth 20x is an empirical question that a few thousand engineering teams are currently answering with their credit cards.

The takeaway

The 46% figure is softer than it reads — one peak week, on one gateway, in the unit most favorable to the claim. But the floor underneath it is hard: five straight months above 30%, up from 4.5% a year and a half ago, on a platform where developers make an unsentimental choice every time they change a config line.

The uncomfortable part for Washington is that export controls appear in this story twice, and both times they help the Chinese models. They constrained American availability in June. They cannot constrain open weights at all. Whatever the policy was designed to do, the token flow suggests it is currently subsidizing the competition it was meant to contain.