Open models with a 99% guarantee: Together AI launches reserved inference

Open AI models now have reliable inference: on July 8, Together AI introduced Provisioned Throughput, with a 99% uptime guarantee and reserved capacity.

What is Provisioned Throughput and how does it work?

Until now, companies working with open-weight models had to choose between two options. The first is serverless inference: convenient and pay-as-you-go, but with no throughput guarantees, which makes it unreliable for production workloads. The second is dedicated GPUs: predictable and fully controllable, but it requires billing in GPU-hours, capacity planning and management of the inference stack. Provisioned Throughput is a third model.

The customer buys PTUs (Provisioned Throughput Units), a fixed share of capacity: a guaranteed number of tokens per minute for a specific model, reserved exclusively for that customer. Input, cached and output tokens draw down PTUs at different rates, which lets teams optimize costs for their actual traffic. The Provisioned Throughput API is identical to the API for Serverless and Dedicated Inference, so customers can switch without changing code.

Key product parameters:

• Launch: July 8, 2026; minimum contract term of one month
• Models available at launch: MiniMax M3 and GLM-5.2
• Regions: North America, EMEA and others
• SLA: 99% uptime
• Price: $0.05 per PTU per minute
• Savings: up to 90% compared with Claude Opus 4.8 at public list prices

Why open models are displacing closed APIs

Together AI reports a sharp rise in demand for open models: in nine months, the volume of tokens through the company's API grew from 30 billion to more than 400 trillion per month. A significant share of that traffic is workloads that previously ran through the closed APIs of the large labs.

Companies that have moved to open models report inference costs falling by 6 to 20 times. The main reason is the ability to distribute tasks along a "Pareto frontier": hard tasks go to powerful models, routine ones to lighter, cheaper models. That flexibility is essential for scaling AI agents: when every employee runs several parallel agents for coding, financial calculations and automation, the token budget becomes a line item set at board level.

Provisioned Throughput addresses exactly this: capacity known in advance, clear pricing and an SLA, without having to hire a DevOps team to maintain GPU clusters.

What this means

Provisioned Throughput closes a gap that has long existed in the open model market: predictable capacity and an SLA while keeping the simplicity of per-token billing. For AI teams scaling agentic workloads, it is a real route from closed APIs to open models without losing reliability.

Frequently asked questions

Which models are available in

Provisioned Throughput right now?

At launch on July 8, 2026, MiniMax M3 and GLM-5.2 are available. Together AI has announced an expansion to 40+ models selected for production use, but has not given a timeline.

How much does the minimum contract cost?

The minimum term is one month. The cost depends on the number of PTUs: $0.05 per PTU per minute. Each PTU guarantees a fixed throughput in tokens per minute for the chosen model, with no additional infrastructure costs.

How can you use open models reliably in production?

Previously, teams had to choose between serverless inference with no throughput guarantees and dedicated GPUs. Provisioned Throughput from Together AI offers a third path: reserved capacity with a 99% uptime guarantee and per-token billing, as with closed APIs.