OpenRouter rolled out US regional routing on September 9. Developers send requests to us.openrouter.ai/api/v1, and those requests are decrypted and processed inside the United States, handed only to inference providers that run there. The same mechanism already exists for the EU, at the endpoint eu.openrouter.ai/api/v1. Both endpoints are limited to the Business and Enterprise plans.
No provider in region, and the request gets a 404
The routing logic is strict: if a model has no available inference provider inside the specified region, the request simply returns a 404 — it will not fall back to sending the data outside the region just to get an answer.
OpenRouter calls this end-to-end regional routing, drawing a line against approaches that only guarantee the inference step stays in-region. The gap is that server-side tool calls under those approaches can still land outside the region, taking the data with them. In a compliance audit, being able to say all processing happened in-country is a stronger commitment than only being able to say inference happened in-country.
Who's on the list
The announcement names Baseten, Fireworks and Azure as US-side inference providers. Models that can run inside the US are listed as DeepSeek V4 Pro, Kimi K3, GLM 5.2, Nemotron 3 Ultra and Inkling. On the EU side, the named model is GLM 5.2, running in Mistral's data centers.
For overall model-vendor coverage, the announcement lists OpenAI, Anthropic, Google, xAI, Meta, NVIDIA, Thinking Machines and Mistral.
The announcement also included a set of share figures: from September 2025 to August 2026, the token share of open-weight models on OpenRouter rose from 26% to 60% in the US and from 16% to 65% in the EU. In one year, open-weight models went from a minority to a majority in both markets.
What it means for the Chinese side
Three models from Chinese teams — DeepSeek V4 Pro, Kimi K3 and GLM 5.2 — appear on the availability list for US regional routing, and the mechanism isn't hard to unpack: they're open-weight, so the weights can be pulled down by US inference providers like Baseten and Fireworks, deployed in US data centers, and served to US enterprise customers with data-residency requirements. Where a model originates and which jurisdiction the data lands in are decoupled along this chain.
For domestic model teams, this route offers a way to reach customers beyond their own API revenue. Without touching customer data, without setting up a US legal entity, and without running their own compliance process, having weights hosted by a third party is enough to get onto a compliance-bound procurement list. The trade-off is that this usage doesn't count toward the team's own API numbers — the model provider can't see who the users are or what they're doing with it. That's a trade-off open weights have always carried, and regional routing just pushes it into a more concrete scenario.
Closed-weight vendors are taking a different path. Companies like OpenAI, Anthropic and Google already have deployment capacity in both the US and the EU, so for them regional routing is more about wiring existing compliance assets into a single API, saving customers the trouble of signing separate data-processing agreements one by one.
What hasn't been disclosed
The announcement gives no pricing details for either endpoint, and doesn't say whether Business and Enterprise differ in how regional routing works for them. How the list of in-region models will expand, and on what timeline, is also left unstated.
For a US enterprise already using OpenRouter, switching over is just a one-line base URL change — provided the model it needs is actually on that list. Calls for anything outside the list will hit the 404, and that's built into the product by design.
Sources: OpenRouter's official announcement, CocoLoop, OpenRouter documentation; endpoint addresses, inference providers, the list of available models and open-weight token share figures verified against the official announcement.