Is the Together AI API reachable right now?

A live probe of api.together.xyz from NetOkay’s edge, and why a model that has not been used recently can look like a network problem.

In short: The card below asks api.together.xyz/v1/models without a key and expects a 401. If the host answers here but your first request to a model waits a long time, that is often a model loading, not your connection; a 30‑second idle limit on your path turns it into a failure.

Last updated 15 September 2026

Live probe · NetOkay Worker → api.together.xyz

Right now, from NetOkay’s edge

Loading the latest observation of api.together.xyz…

One keyless GET to https://api.together.xyz/v1/models, cached for a minute per Cloudflare location. On a healthy day this probe sees HTTP 401, the API asking for a key.

Run the full website check →Test this browser →Compare environments →

What this probe shows, and what it does not

The NetOkay Worker resolves api.together.xyz through Cloudflare DNS, opens a TLS connection and sends one GET without credentials, following no redirects. Any HTTP status, including 401, 403 or 405, means the host answered. Several providers refuse keyless automated requests at their edge with a 403 in a few milliseconds; that still proves DNS, TLS and the front door, but it is the provider’s edge answering, not necessarily the API application behind it. The result is one observation from one Cloudflare location, refreshed at most once a minute and never stored.

It cannot see your API key, quota, region, model availability or account state, and it does not travel your network path. A healthy probe and a failing application are compatible; the sections below are for that case.

Together AI specifics

API host: api.together.xyz, HTTPS on port 443, OpenAI‑compatible under /v1/.
Streaming: Server-sent events with data: JSON chunks and data: [DONE]. Together also supports its own /inference shape, but OpenAI clients use the /v1/ surface unchanged.
Model start‑up: Serverless endpoints for less frequently used open models may need to load before the first token. That first‑byte wait can exceed idle limits on corporate proxies; the stream itself, once started, is fast.
Error codes: 401 invalid key; 429 rate limit; 503 or a JSON error when a model is unavailable or still loading; 400 unknown model name.
Official status: status.together.ai lists incidents by service.

Reachable here but failing for you?

Run the homepage checks from the network where the app fails. The 30‑second idle stream shows whether something on your path closes silent connections, which is what a long model pause looks like to a proxy.
Read the idle result. Closed during the silence points at a proxy, gateway or NAT idle limit under 30 seconds. Start and done arriving together points at response buffering. Both have fixes in the streaming guide.
Compare environments. If the browser passes and a container, server or agent fails, compare the public IP each one uses, then run the CLI from the failing environment against api.together.xyz.

Together AI API status FAQ

The first request to a model takes ages, then it is fast. Why?

The model was loading. This is a first‑byte delay rather than a streaming problem. If your client or proxy times out during that wait, raise the timeout or send a keep‑alive request first.

The probe is healthy but my dedicated endpoint fails. Does this page cover it?

No. Dedicated endpoints have their own hostnames. This page only observes the shared api.together.xyz host.

Does a 401 here mean Together is fine?

It means DNS, TLS and HTTP all work from Cloudflare’s edge to api.together.xyz and the API answered. It says nothing about model availability or your own network.