Together AI Production Checklist

Look forServerless bills per input and output token with no minimum cost and is best-effort. Provisioned throughput reserves capacity for five listed models at US$0.05 per PTU per minute on a one-month minimum, through sales. Dedicated endpoints serve a model on reserved hardware billed by the minute per replica, an H100 80GB at US$5.49 per hour.

Look forRun your tool definitions, structured output and longest prompts through the client you ship. Practitioners report tool calls failing on gpt-oss-120b through ChatTogether and on Qwen3-Coder-Next through LiteLLM, and a grammar error on Llama 3.3 70B. Confirm each model is in the serverless catalog.

Look forA developer reports notice emails from Together deprecating GLM-5.1 on July 10, 2026, with no upgrade path. A models.dev issue found six models still listed as active that the serverless endpoint no longer serves. MLCommons' evaluation broke twice when annotator models left serverless. Keep a tested replacement for every pinned model, or serve it on a dedicated endpoint.

Look forAccess needs a minimum US$15 credit purchase and is suspended at zero, and the cut-off is not an exact cap. API keys have no spending limits of their own. Expect 429 or 503 responses at high demand, and retry with backoff.

Look forTogether stores prompts and responses by default, and admins turn storage off for zero data retention. Passthrough models send data to the upstream provider under its policy, behind a separate organization toggle. US-only processing covers where requests are processed, and data storage falls outside that scope.

Look forSSO via SAML and OIDC is available for Scale and Enterprise accounts, and SCIM is listed under What's coming. Audit and lifecycle events are listed per dedicated endpoint, so record organization changes in your own change log.