Summary: Together AI’s serverless docs now document a US-only endpoint, us-api.together.ai, that keeps inference processing inside the United States with no GPUs to provision. You point your client at the new base URL, keep your existing API key and model IDs, and pay a 1.25× premium on token pricing. For builders serving regulated US customers on serverless models, this changes what you can promise in a contract — with some limits worth reading carefully.
What the US endpoint actually guarantees
The scope is narrow and precise: the endpoint covers where your request is processed. Per the Together AI docs, requests to us-api.together.ai keep inference processing in the US, while requests to the standard api.together.ai do not guarantee US processing.
That means the guarantee does not cover where data is stored, where logs are written, or how network traffic is routed. If your compliance requirement is really about data residency end to end, this endpoint alone won’t close that gap — the docs point you to their privacy and security options for private networking or stricter residency needs.
How you use it
The integration is a base URL swap. Same SDK, same key, same model strings:
import Together from "together-ai";
const client = new Together({ baseURL: "https://us-api.together.ai" });
const response = await client.chat.completions.create({
model: "zai-org/GLM-5.3",
messages: [{ role: "user", content: "Hello!" }],
});Or from a shell:
curl https://us-api.together.ai/v1/chat/completions \
-H "Authorization: Bearer $TOGETHER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "zai-org/GLM-5.3", "messages": [{"role": "user", "content": "Hello!"}]}'The supported model list is open-weight heavy: Kimi-K3, GLM-5.3 and GLM-5.3-Flash, several DeepSeek-V4 variants, and MiniMax-M3.1-Flash. That fits the broader pattern I covered in /blog/muse-spark-open-together-ai-api-pricing-en/ — Together’s serverless catalog is where open models tend to show up first.
The 25% premium and the 404 behavior
Two operational details deserve attention before you route production traffic.
First, pricing: US-only inference costs 1.25× the standard serverless rate across input, cached input, and output tokens. If your margin model is tight, that premium applies to every token, including cached ones — worth modeling before you commit.
Second, failure semantics: there is no fallback outside the selected geography. If a model isn’t available there, the request fails with a 404 of type region_unavailable rather than silently running somewhere else. That’s the right design for compliance — silent fallback would defeat the point — but it means your retry logic needs to target the same geography-specific endpoint. A retry that falls back to api.together.ai would quietly break your guarantee.
Where this fits
This sits in the middle of a spectrum I’ve written about before: an earlier post on in-region routing distinguished merely routing encrypted traffic into a region from actually running inference there. Together’s endpoint is the latter — processing happens in-geography — but stops short of the confidential-computing story where even the host environment is attested.
The practical takeaway: if your requirement is “don’t process customer prompts outside the US,” this is a one-line change and a 25% surcharge. If your requirement is full data residency covering storage, logs, and network paths, treat this as a partial answer and go through Together’s privacy and security options next. Verify the endpoint behaves as documented in a staging environment before you write it into a customer commitment.
Sources
View the original post on X