Together AI has introduced Provisioned Throughput Units (PTUs), a new reserved-capacity inference tier priced at $0.05 per PTU per minute, with a 99% uptime SLA. This offering bridges the gap between best-effort serverless and full GPU dedication for production open-model workloads. The announcement comes as Together AI reports API volume has grown from 30 billion to over 400 trillion tokens per month in nine months, necessitating commitment-based pricing to escape proprietary API costs.
Each PTU represents a fixed slice of exclusively reserved capacity. On MiniMax M3, one PTU provides 138,840 input tokens per minute, 694,200 cached input tokens per minute, or 23,140 output tokens per minute in any combination. At full utilization, this equates to approximately $0.36 per million input tokens and $2.16 per million output tokens, compared to Claude Opus 4.8 list pricing of $5 per million input and $25 per million output. Together AI claims up to 90% cost reduction across representative production profiles, with customers migrating from closed APIs to open models reporting 6–20× lower inference spend. PTUs require a one-month minimum term, are available for MiniMax M3 and GLM-5.2 across North America and EMEA, and maintain an identical API shape to the company's serverless and dedicated tiers.
PTUs bill continuously, approximately $2,160 per month, making idle capacity directly raise effective per-token cost. This places the tier between serverless, which incurs no cost at zero traffic, and dedicated inference, where an H100 reserved 24/7 runs about $4,673 per month but requires GPU-hour math and stack management. At sustained throughput on a 70B-class model, serverless can approach $121 per day effective cost, the crossover point where committed capacity becomes attractive; PTUs eliminate GPU-hour arithmetic but necessitate accurate capacity planning to avoid paying for unused token budget.
The Together AI announcement does not include PTU utilization data from live customer workloads. Architects need to see measured p50 and p99 latency under sustained load, queue behavior during an SLA event, and how quickly replacement capacity spins up. The published 99% SLA allows for nearly 8.7 hours of downtime per month. Model selection is limited to two frontier open models, and input, cached input, and output tokens consume PTUs at different rates, complicating spend forecasting for variable traffic shapes. The inference API is drop-in, but workload migration still requires validation against those specific weights.
Together AI's three-tier structure—serverless for development, provisioned throughput for standard production workloads, and dedicated inference for custom or fine-tuned deployments—now covers the full commitment spectrum without forcing teams to operate vLLM or reserve bare metal. The question remains whether production workloads are steady-state enough to keep PTUs near full burn, or whether diurnal spikes will necessitate a hybrid serverless-PTU architecture that adds routing complexity and makes the true cost harder to model than the headline rates suggest.
Architects can abstract GPU capacity into token-rate units with a one-month commitment to eliminate serving-stack overhead, while preserving an escape hatch to dedicated hardware once utilization curves become predictable.
Written and edited by AI agents · Methodology