The number is 46%. That is the peak weekly share of enterprise token volume on OpenRouter, as of mid-2026, that ran on Chinese-origin AI models. Every week since February 8 it has been at least 30%. Twelve months ago it averaged 11%. Eighteen months ago, in the first half of 2025, it was 4.5%.
That is a category-shape change, not a rounding change, and the shape it changed to is worth being precise about. It is not that enterprises decided the frontier moved to Hangzhou. It did not. Anthropic still writes the top of the reasoning curve on hard agentic tasks. OpenAI still owns the top of the government-vetted list and most of the consumer surface. Google is still, for anything involving a search index or a browser context, structurally advantaged. The frontier is where it was. What changed is that the frontier stopped being the thing that mattered for most tokens.
The router replaced the pick
For approximately eighteen months, the running assumption in enterprise AI adoption was that you picked a model. Sometimes you picked two, one for coding and one for chat, and called that your “AI strategy.” The buying motion was model-selection, singular. The vendor pitch was model-quality, singular. Every RFP had a chart that ranked the models against each other on a set of benchmarks, and the winning model got the traffic.
That is not how the tokens flow anymore. The operating pattern is a router in front of a portfolio, and the router picks per-request. Simple queries route to DeepSeek V4 Flash at $0.14 per million input tokens. Reasoning queries route to Opus 4.8 or GPT-5.6 Terra at fifty times that price. Coding queries route to Grok 4.5 or Sonnet 5 depending on which is cheaper for the day’s context length. Data-extraction queries route to Qwen. Vision queries route to whichever Gemini variant has the latest quality-per-dollar ratio in the OpenRouter dashboard, which updates weekly.
Nobody sat down and decided this. Nobody wrote a strategy doc. It grew, because the price gap is that stark. DeepSeek V4 Flash: $0.14 per million input tokens. GPT-5.5: $5.00. That is not a 10% delta a procurement team can wave away. It is a 35x delta that a competent engineering team is embarrassed not to be capturing on the tasks where it does not matter which model runs. And on any given enterprise workload, the honest answer to “does it matter which model runs” is: for most of the tokens, no.
What “tokenmaxxing” was, and why it died
The term “tokenmaxxing” was a mostly-affectionate description of what a lot of teams were doing in 2024 and 2025. Route everything to the best available model, spend the money, worry about margins later. The assumption behind it was that model quality was the bottleneck, that the highest-quality model would always produce enough marginal value on any given task to justify the marginal cost, and that anyway the token prices would keep falling on the frontier faster than costs would grow on the volume side. So tokenmaxx. Ship the flagship. Explain the AWS bill to the CFO later.
Three things killed it. First, the frontier prices did not fall fast enough. GPT-5.6 pricing is not meaningfully cheaper than GPT-5 pricing, per capability. Opus 4.8 is not meaningfully cheaper than Opus 4.5. The flagship line got better without getting proportionally cheaper, which broke the tokenmaxxing math on the second derivative.
Second, the non-frontier prices fell off a cliff. DeepSeek, Qwen, Kimi, and the other open-weights Chinese lines cut compute costs through architectural work, mixture-of-experts routing, aggressive quantization, and (per SemiAnalysis) a willingness to run inference at close to marginal cost while the parent companies figured out what the actual business was. What used to be a 3x-5x price gap between the frontier and the good-enough tier turned into a 30x-50x gap.
Third, the capability gap on most workloads is now smaller than the price gap. This is the specific thing that changed in 2026. For summarization, extraction, classification, translation, structured-output generation, RAG-style question-answering over a bounded corpus, and a large chunk of the routine chat-assistant surface area, DeepSeek V4 or Qwen 3.5 produce output that is empirically hard to distinguish from Opus in blind eval. Not on hard reasoning. Not on long-horizon agentic tasks. But on most of the tokens most enterprises are actually spending, yes.
Once those three things were true simultaneously, tokenmaxxing became strictly worse economics than routing. The market noticed. The market moved.
Where this leaves the frontier labs
Anthropic and OpenAI are not in trouble. They are, however, in a different business than they thought they were in eighteen months ago. Watch what they are doing, not what they are saying.
OpenAI’s Deployment Company (per the July 8 Northslope acquisition) is buying forward-deployed engineers who can implement systems inside customer offices. That is not a bet on winning the tokens. That is a bet on capturing the implementation dollars, because the tokens are commoditizing under them. Anthropic’s Claude Corps (per this week’s application deadline) is spending $150 million to place trained fellows inside 1,000 nonprofits. That is a workforce-development play and a distribution play. It is not a token-share play. Both companies are behaving like the token wars are decided and they lost the low end and are pivoting to services, deployment, and enterprise-of-record status where the seat prices are stable and the margins are on the human side of the transaction.
The frontier itself still matters, but it matters differently. It is the ceiling that anchors what the whole market will pay for a token. When Opus 4.8 or GPT-5.6 Terra defines what “best possible answer” looks like on a hard task, the router uses that as the standard, and everything below it prices as a percentage discount off that ceiling. The frontier labs are getting paid to be the ceiling. They are increasingly not getting paid to be most of the volume.
What the routing layer looks like a year from now
Three things are already visible.
One, the router itself becomes a real product category. Not a middleware wrapper. A proper layer with its own cost accounting, quality evals, model-selection logic, fallback trees, and prompt-caching strategy per destination model. OpenRouter is the obvious current example. Portkey, Vercel AI Gateway, Braintrust, and every major cloud’s inference-router are the enterprise-flavored versions. Somebody wins this layer. Whoever wins it captures a fraction of every enterprise AI dollar spent for the next decade.
Two, the model providers push back against the router by trying to lock context to their surface. This is why Anthropic launched skills. It is why OpenAI is pushing memory + workflows into ChatGPT Enterprise. It is why Google is trying to make Gemini the default inside Workspace. If your context lives inside my platform, the router at the front of the request cannot cleanly send the query somewhere else, because the somewhere-else does not have your history. The lock-in war moves from model quality to context gravity.
Three, the Chinese models keep getting cheaper faster than the American ones. This is not ideological. It is the shape of the compute markets, the shape of the labor markets, and the shape of the willingness to run inference near cost for strategic reasons that a US public company cannot copy. The 46% number is not the top. It is not obvious where the top is. It probably is not 50/50. But the direction of travel does not have a natural stopping point at any given US-lab market-share number.
The tokenmaxxing era looked like a competition between model brands. The routing era looks like a competition between infrastructure layers. That is a boring story for the trade press and a fascinating one for the CFO. Which, if you are paying attention, is the actual signal.