Together AI Review: Open Models Billed per Token, With Tool Calls and Model Retirements to Test

Limits and billing

• Rate limits: Together's serverless performance is best-effort. When demand is high, Together may limit requests, and you may receive 429 Too Many Requests or 503 Service Unavailable responses.
• Contracts: purchased rate limits under an active Enterprise or Scale contract stay in place until the contract expires, and when a model is in especially high demand Together may apply custom rate limits or access restrictions to that model.
• Batch: batch jobs run against a separate rate-limit pool, with up to 50,000 requests per batch, up to 100 MB per input file, up to 10 MB per line, which inline base64 images count toward, and up to 30B tokens enqueued per model at any time. The completion window defaults to 24h, cannot be changed and is a best-effort target.
• Spend: the zero-balance cut-off does not guarantee an exact spending cap. Metering has a short delay, so some usage can arrive after the balance reaches zero and the customer remains responsible for it. API keys do not currently support independent spending limits, and organization and project budgets and spend alerts are planned features that are not currently available to configure.
• Provisioned throughput: PTUs reserve capacity for the contract term. Unused PTUs don't roll over and aren't refunded or credited, and overage traffic is not covered by the provisioned throughput SLA.

Models

• Together Link is currently in beta. By default, its sessions use the auto router, which picks a Together AI model for each request and, in Claude sessions with an Anthropic API key, sends the most difficult requests to Claude Opus. To choose a model, Together Link lists available models, context limits and pricing, and a session is pinned to a single model by passing --main with the model ID before the tool name. Together Link refuses or ignores a model flag placed after the tool name.
• Catalog: Together's serverless catalog lists chat models by API model string, including zai-org/GLM-5.3, openai/gpt-oss-120b and moonshotai/Kimi-K3. Serverless and dedicated model inference support different sets of models.
• Provisioned throughput covers GLM-5.3, GLM-5.3 Flash, DeepSeek V4.1 Flash, GLM-5.2 and Kimi K3, with a reliability commitment that at least 99% of eligible requests complete successfully each month.

Data, security and sign-in

• Storage: by default Together stores prompts and responses and may use them for product improvements. Organization admins can turn storage off to enable zero data retention. Data sharing for training other models is opt-in and not enabled by default.
• Passthrough: some Together models are offered as passthrough. Together forwards prompts and responses directly to the upstream provider, and data is handled under that provider's own data policy. A separate organization-level toggle controls passthrough models, independent of the training opt-in.
• Certifications: the trust center lists SOC 2 and ISO 27001:2022 under Compliance, with a SOC 2 Type 2 Report and an ISO 27001:2022 Certificate among its resources. Together says it completed a SOC 2 Type 2 examination, an independent audit spanning several months (Jul 8, 2025).
• SSO: Together supports SSO via SAML and OIDC with Google Workspace, Okta, Microsoft Entra and JumpCloud, for Scale and Enterprise accounts. SCIM provisioning is listed under What's coming.
• Audit: for dedicated endpoints, Together lists an endpoint's audit and lifecycle events newest first, merging endpoint events with those of every deployment under the endpoint, and can filter to one subject, such as a rollout. None of the 417 pages in Together's documentation index on Nov 1, 2026 uses the term audit log.

Region and integrations

• Region: on Together's geo-specific endpoint there is no fallback outside the selected geography. If a model isn't available there, the request fails with a 404 error of type region_unavailable.
• Hosting: models from third-party authors such as DeepSeek, Qwen and Mistral that Together hosts run on Together's own infrastructure and do not call out to the model author. DeepSeek models are hosted in Together's North America data centers.
• SDKs: Together's API is compatible with the OpenAI REST API and SDKs. The Vercel AI SDK's @ai-sdk/togetherai provider, LiteLLM 1.100.0 and later, and LangChain's langchain-together package, for chat models and embeddings, support Together.