Together AI US-Only Inference: Pricing, Endpoint and Fallback Rules

Together AI now documents US-only serverless inference at https://us-api.together.ai/v1, with a 25% premium on input, cached input and output tokens. The route uses existing API keys and model IDs. Its guarantee covers inference processing, not storage, logs or network routing.

For teams evaluating open models against a geographic processing requirement, that is a useful option. It also creates a straightforward engineering question: can your application keep using the regional route when something fails, rather than quietly retrying through its ordinary endpoint?

Which models support the US endpoint?

Together's geo-specific inference documentation lists seven supported model IDs. The two DeepSeek Flash entries below are separate catalog choices, not two additional model families.

Model ID	Input	Cached input	Output
MiniMaxAI/MiniMax-M3.1-Flash	$0.1875	$0.01875	$0.75
deepseek-ai/DeepSeek-V4-Pro-0813	$1.65	$0.1625	$4.95
deepseek-ai/DeepSeek-V4.1-Flash-1017	$0.375	$0.0075	$1.50
deepseek-ai/DeepSeek-V4.1-Flash	$0.375	$0.0075	$1.50
zai-org/GLM-5.3-Flash	$0.1875	$0.0375	$0.625
moonshotai/Kimi-K3	$3.75	$0.375	$18.75
zai-org/GLM-5.3	$1.75	$0.325	$5.50

These are US dollars per million tokens from Together's pricing page, checked November 12. Do not substitute a similarly named newer checkpoint and assume it has the same regional availability.

Calculate the premium on your actual traffic

Consider a hypothetical GLM-5.3 workload with 10 million uncached input tokens and 2 million output tokens. At standard rates of $1.40 input and $4.40 output, that is $22.80. At the US-only rates above, it becomes $28.50: a $5.70 difference for the same token mix.

That calculation excludes cache hits and any other application costs. A realistic forecast should separate uncached input, cached input and output, then apply each rate. Do not count a cached token twice as both ordinary input and discounted input.

The premium may be a small part of an agent's total cost, especially when the agent invokes paid tools. Nevertheless, a geography requirement belongs in the routing budget from the beginning. Comparing standard-price model A with region-constrained model B can make a model-selection spreadsheet look more decisive than the actual deployment choices allow.

Point the request at the regional host

For an already authorized account, the documented HTTP pattern is:

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": "Summarize this synthetic support ticket: password reset requested."}]
  }'

This example has not been executed here. Start with synthetic content and inspect the application's configured destination before introducing customer data. In a shared gateway, check the route used by background jobs as well as interactive requests; they may use different client configurations.

A regional failure should stay a regional failure

Together documents a 404 error with type region_unavailable when a model is unavailable in the selected geography. It does not automatically run that request elsewhere. The standard api.together.ai endpoint does not guarantee US processing.

The dangerous fallback can therefore live in your own application. A generic retry handler may switch hosts, select another provider or resend through a global gateway after the regional call fails. If the geographic boundary is mandatory, test those branches explicitly and keep retries within the permitted route.

Use an artificial failure in a test harness and observe the destination of the next attempted call. This checks your routing policy without pretending that a successful model response proves physical processing location.

Processing location is only one part of the review

A US inference call can still sit inside a workflow that stores prompts, traces and outputs somewhere else. Map those components separately and ask for the applicable storage and contractual commitments rather than extending the inference promise yourself.

The useful change is a serverless processing option with a visible price. Treat it as one constraint your architecture can satisfy, then verify the surrounding system with the same care.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn infrastructure choices and product constraints into clear stories developers and customers can use.

Book a 30-minute call