The following are selected Together AI list-price observations checked 26 October 2026, not a complete market price range, negotiated quote or quality-equivalent comparison. Prices can change; taxes, capacity availability, commercial terms and other services require confirmation.
| Service/model | Published unit | Observed price | Caveat |
|---|---|---|---|
| Qwen2.5 7B Instruct Turbo serverless inference | Per million input / output tokens | $0.30 / $0.30 | Does not price a custom-model project or human review |
| Llama 3.3 70B serverless inference | Per million input / output tokens | $1.04 / $1.04 | Different model and quality profile; not interchangeable by price alone |
| Llama 3.1 8B LoRA supervised fine-tuning | Per million processed tokens | $0.34; $4 minimum/job | Training tokens include epochs and applicable evaluation processing |
| Llama 3.1 8B full supervised fine-tuning | Per million processed tokens | $0.38; $4 minimum/job | Serving and engineering are separate |
| H100 dedicated inference | Per GPU-hour, on demand | $5.49 | Different service from cluster capacity |
| H100 GPU cluster | Per GPU-hour, on demand | $3.99 | Not a production-service quote |
Source and exact billing definition: Together AI pricing. The fine-tuning table distinguishes LoRA and full fine-tuning; reproduce that distinction when seeking a quote. No claim is made that these are the cheapest available services.
How do we turn those prices into useful arithmetic?
Illustration, not a performance result: 50,000 requests with 6,000 input and 1,000 output tokens each generate 300 million input and 50 million output tokens. At the selected $0.30/$0.30 rate, model usage is $105. That excludes retries, parsing, tools, storage, review and operating labour; it says nothing about whether the outputs are acceptable.
For training, a 10-million-token dataset processed for three epochs gives 30 million training tokens before any billable validation. At $0.34 per million, that component is $10.20, above the cited $4 minimum. Cheap metered training does not price dataset creation, experimentation or production integration.
For capacity, a hypothetical 730 billed hours at $5.49 gives $4,007.70 for one GPU. That is a budgeting illustration, not a claim that one GPU can serve the buyer’s workload or that every month has 730 hours. Measure throughput, headroom and redundancy before translating capacity into tasks.