Together AI puts a fixed price on reserved open-model capacity with Provisioned Throughput
Together AI's Provisioned Throughput sells reserved inference capacity for open models by the minute, with a 99% uptime SLA and predictable cost. It is a bet that teams running open weights in production want a bill they can forecast, not just a cheap per-token rate.
Together AI has launched Provisioned Throughput, a way to reserve a fixed slice of inference capacity for frontier open models and pay for it by the minute rather than by the token. The offering carries a 99% uptime commitment and a predictable monthly bill, and it targets the exact moment a team stops prototyping on open weights and starts depending on them. The pitch is less about a lower headline price than about a cost you can put in a spreadsheet before the traffic arrives.
What shipped
The unit of purchase is a Provisioned Throughput Unit, or PTU, priced at $0.05 per PTU per minute and held exclusively for the buyer. On MiniMax M3, one PTU delivers a guaranteed 138,840 input tokens per minute, 694,200 cached input tokens per minute, and 23,140 output tokens per minute, or any mix within that envelope. At full utilization that works out to roughly $0.36 per million input tokens and $2.16 per million output tokens, which Together frames as up to 90% cheaper than comparable proprietary APIs. Reserved capacity is available today for MiniMax M3 and GLM-5.2, with more models promised.
The commercial shape matters as much as the numbers. Provisioned Throughput comes with a one-month minimum term, volume discounts at higher commitments, and capacity in North America, EMEA, and beyond. That is a deliberate contrast with Together’s serverless tier, which bills per token on best-effort capacity. Serverless is the right default when traffic is spiky or unknown. Reserved capacity is what you buy when a product depends on inference being there at a known rate, and when finance wants the number to stop moving.
For a working team, the practical read is simple: this is the tier you graduate to when an open model is no longer an experiment. It trades the pay-only-for-what-you-use appeal of serverless for a guaranteed rate and a forecastable invoice.
Where this lands in the market
Provisioned throughput is not a new idea. Azure OpenAI popularized the PTU acronym, and dedicated-capacity tiers exist across the hosted-inference field. What is notable is Together attaching that enterprise-grade purchasing model to open weights specifically. The company raised $800M earlier this month on the thesis that demand for open-model inference is compounding, and this launch is the productized version of that argument: if teams are going to run MiniMax and GLM in production, they need the same reliability guarantees they would expect from a closed API.
The timing rhymes with a broader shift. Hugging Face’s CEO argued this week that companies are increasingly done renting their AI and want to own more of the stack, and reserved open-model capacity sits squarely in that move. Owning the weights removes vendor lock-in on the model; reserving the throughput removes the unpredictability that usually pushes risk-averse teams back toward a managed proprietary endpoint. Together appears to be betting that the last real objection to open models in production is operational, not qualitative, and pricing the objection away is the play.
What’s worth watching
The economics only hold if utilization holds. A PTU is cheapest per token at full saturation, which means the savings depend on steady, high-volume traffic. Teams with bursty or seasonal load could end up paying for reserved capacity they do not fully use, so the real comparison is not PTU price versus serverless price but PTU price at realistic utilization versus a per-token bill. That math will decide whether this is a discount or a commitment trap for any given workload.
The other signal is model coverage. Two models at launch is a start, and the value grows sharply as the reserved-capacity catalog widens toward the models teams actually want to standardize on. Over the next 90 days, the things to watch are how quickly Together adds models, whether competing open-inference providers answer with their own reserved tiers, and whether published case studies show the utilization levels that make the pricing work. Stackmaven’s follow-up coverage will land on or around October 10.
- Open, convenient and predictable: Introducing Provisioned Throughput (Together AI) www.together.ai
- Provisioned Throughput redefines open-model inference economics and predictability (Futurum Group) futurumgroup.com
- Together AI pricing www.together.ai