Together AI launches Provisioned Throughput for reserved open-model inference at 90% below closed APIs

Together AI introduced Provisioned Throughput, a reserved inference capacity product for frontier open models (MiniMax M3, GLM-5.2 initially) priced in Provisioned Throughput Units (PTUs) at $0.05 per PTU per minute. Each PTU delivers a guaranteed fixed rate of tokens per minute—e.g., on MiniMax M3: 138,840 input tokens/min, 694,200 cached input tokens/min, or 23,140 output tokens/min—with a 99% uptime SLA and one-month minimum commitment.

At full utilization, a PTU on MiniMax M3 works out to roughly $0.36 per million input tokens and $2.16 per million output tokens against $5 and $25 list prices on Claude Opus 4.8, representing ~90% cost savings. Unlike Together's serverless API (pay-per-token, best-effort), Provisioned Throughput gives teams "pay for guaranteed capacity" with predictable pricing that scales with their traffic shape. No GPU-hour math or infrastructure management—the same token-based consumption model enterprises expect from proprietary providers, pointed at open models.

Together's token volume has grown from 30 billion to over 400 trillion tokens per month in nine months, with a substantial share now coming from workloads that migrated off closed APIs. Provisioned Throughput addresses the production-readiness gap: serverless is best-effort (fine for dev), dedicated GPU endpoints require full complexity management. This middle tier offers reliability guarantees without operational overhead.

For architects: Open models now occupy a growing fraction of the Pareto frontier—coding, finance, marketing, and workflow automation tasks now run 6–20x cheaper on open alternatives than proprietary APIs. Provisioned Throughput makes the migration path from closed to open viable for production: frontier-quality models, token-based pricing you can budget, and SLA guarantees your customers depend on, at a tenth the cost.