US Models Fell From 70 to 30 Percent of OpenRouter Traffic in a Year

American models went from 70 percent of OpenRouter token traffic to about 30 percent in twelve months. Chinese open-weight models took the rest.

The driver is not benchmarks. It is a 60 to 90 percent price gap, and it is finally big enough that engineers are routing production workloads across it.

Coding was the wedge. Two Chinese models now handle roughly half of all coding tokens on OpenRouter, because free previews and 1M-context windows made them the default inside AI IDEs.

OpenRouter sits in front of hundreds of models and routes real API calls from real applications. It is one of the cleaner signals we have for what developers actually ship on, not what they say in a survey. And the signal over the last year is stark. A year ago, US models were about 70 percent of the tokens flowing through it. Now they are around 30 percent. Chinese open-weight models, which were under 2 percent eighteen months ago, crossed 45 percent in the spring and have been reported north of 60 percent of token consumption by mid-year.

DeepSeek alone commands roughly 16 percent of all token volume, more than any single provider from Google, Anthropic, or OpenAI. That is one Chinese lab outweighing each of the American frontier labs on the busiest model router in the world.

This is a cost story, not a quality story

The easy read is that Chinese models got good. They did, but that is not what moved the traffic. What moved it is that they got good enough at a price that is not close.

OpenRouter's own people have put the gap at 60 to 90 percent cheaper for Chinese open-weight models versus the Western frontier. DeepSeek's cheaper tiers run well under a dollar per million tokens on both input and output. Anthropic and OpenAI flagship pricing sits at multiples of that, and the recent moves have been up, not down. Sonnet's intro pricing already expired into a higher band. Fable 5 shifted from subscription-included to metered credits at roughly double the Opus rate.

When your inference bill is a rounding error, none of this matters and you should stay on the best model you can get. When inference is a real line item, a 5x to 10x unit-cost difference stops being a spreadsheet curiosity and becomes the whole conversation. That is the threshold a lot of companies crossed this year, because agentic workloads multiply token counts. An agent that reads, plans, calls tools, and retries burns ten to fifty times the tokens of a single chat completion. The cost gap does not add up linearly. It compounds with every agent step.

The switching is already happening at real companies

This is not developers kicking tires on a weekend. Lindy, the agent-automation company, moved all of its traffic off Anthropic's Claude to DeepSeek. CEO Flo Crivello said the switch saves the company millions. Ramp's spending data, which tracks what businesses actually pay for rather than what they post about, showed DeepSeek leading the foundational-model category by real dollars routed to its API.

Coding is where it started, and the numbers there are the loudest. MiMo-V2-Pro and Qwen 3.6 Plus together account for close to half of all coding tokens on OpenRouter. The playbook was simple: ship a free preview, post strong SWE-Bench numbers, offer a million-token context window, and let the AI IDE ecosystem wire you in as a default. Once a coding agent is pointed at a cheap model that clears the bar, the traffic follows and it does not come back easily.

The catch nobody should skip

I am not telling you to move your production stack to a Chinese model this week. There are three real problems, and the cost savings do not erase them.

Data residency and governance. Sending your prompts, and whatever context they carry, to inference endpoints under a different regulatory regime is a decision your security and legal teams own, not one an engineer makes to shave a bill. For regulated data, that alone can end the conversation. Self-hosting the open weights removes the endpoint problem but adds the cost and skill of running the infrastructure yourself.

Content and behavior constraints. These models carry restrictions on certain topics, and for some applications that shows up as refusals or oddly shaped outputs in places you did not expect. Test against your actual use case, not a generic benchmark.

Stickiness cuts both ways. The same reason coding traffic moved to Chinese models is the reason it is hard to move back. Once you build retries, evals, prompt formats, and tool schemas around a model's quirks, switching again has a real cost. Choose deliberately, because your second migration is as expensive as your first.

What to actually do

The mistake is treating this as a binary. It is not one model or the other. It is a routing decision, per task, and the teams pulling ahead already run it that way.

Split your workload by what the task is worth. A high-stakes reasoning step, a customer-facing generation, a legal summary, those can justify the expensive frontier model. A bulk classification job, a first-pass draft an agent will revise anyway, a retrieval reranker, an internal tool that never touches a customer, those are exactly where a model that is 80 percent cheaper and clears the bar wins. You do not need your best model for every token. You need the cheapest model that passes your eval for each task.

Which means you need the eval first. The reason most teams overpay is that they never built a way to measure whether a cheaper model clears their bar, so they default to the expensive one out of caution. Build a small, honest test set for each task that matters. Then a cheaper model is not a gamble, it is a measurement. Route the tokens the measurement approves and keep the frontier model for the ones it does not.

The 70-to-30 collapse is what happens when a price gap gets wide enough that measurement beats habit. The American labs are not losing on capability. They are losing the tokens where capability was never the thing that mattered.