Together AI as cloud orchestrator - Per-GPU-hour, per-token, and PTU billing models

Fetched 2026-10-04. Together AI offers three distinct billing models. Key structural facts: (1) Serverless Inference: priced per 1M tokens (input/output), with batch API pricing at a discount. This is the most common model for LLM inference. (2) Provisioned Throughput: priced per PTU/min ($0.05/PTU/min), where PTU is a model-specific throughput unit. Each model has different TPM/PTU ratios. (3) GPU Clusters: priced per GPU-hour, with on-demand, preemptible, and reserved tiers. H100 $3.99 on-demand, $1.99 preemptible, $3.69-$3.19 reserved. H200 $5.99/$2.99/$4.99-$3.99. B200 $8.19/$4.09/$7.99-$6.79. B300 $9.99/$4.99/contact. (4) Dedicated Inference: single-tenant, per GPU-hour. H100 $5.49, B200 $8.99. Others contact-sales. Together's GPU Clusters preemptible H100 at $1.99/GPU-hr is the cheapest published managed-provider H100 rate found in this research. The three billing models serve different use cases: serverless for variable workloads, provisioned for predictable high-throughput, and GPU clusters for custom workloads.