Together AI Official Documentation Snapshot
Verified against Together AI primary documentation and the official togethercomputer/skills repository on 2026-09-10. Recheck the live model catalog, pricing, rate-limit headers, and deprecation page before changing a production workload; those surfaces are intentionally dynamic.
SDK and authentication
- Quickstart
- Project-scoped API keys
- Python SDK v2 migration
- Official Python v2 client
- Official Together agent skills
Inference and models
Asynchronous and managed workloads
- Batch overview
- Batch tutorial
- Fine-tuning CLI and lifecycle
- Dedicated Model Inference
- Migrate dedicated endpoints from v1
Operations
- Dynamic serverless rate limits
- Usage and cost analytics
- Error codes
- Model lifecycle and deprecations
Pinned interpretation
- New Python work targets
together>=2.0.0; v1 is maintenance-only and new features land in v2. - SDKs read
TOGETHER_API_KEY; REST requests useAuthorization: Bearer <token>. Keys are project-scoped and must remain in an approved secret store. - Serverless limits are dynamic per organization and model. Read response headers and adapt; do not encode a permanent RPM or TPM table.
- Batch input is JSONL with stable
custom_idvalues. Upload withpurpose="batch-api", reconcile output and error files bycustom_id, and do not assume output order. - Batch and fine-tuning are asynchronous job APIs. Poll boundedly and persist job identifiers; Together does not document a general signed webhook delivery surface for these jobs.
- Current Dedicated Model Inference uses the v2 beta endpoint/deployment resource model. Do not create new capacity through the retired v1 endpoint-create flow.
- Model availability, redirects, context limits, prices, and batch eligibility change. Resolve them from the current catalog and deprecation pages at execution time.