Together AI pricing guide details
Pricing Overview
Together AI is pay as you go, with a separate rate card for each product. Serverless inference is priced per 1M tokens by model, with separate input and output rates. Paid chat models range from $0.04 input (Tev1 4B Experimental) to $3.00 input and $15.00 output (Kimi K3), so the model you pick sets your bill more than anything else.
Capacity is priced by the hour. Dedicated inference on demand costs $5.49 per GPU per hour for NVIDIA HGX H100 and $8.99 for HGX B200. GPU clusters on demand cost $3.99 per GPU-hour for HGX H100, $5.99 for HGX H200 and $8.19 for HGX B200. Other GPUs and reserved capacity are quoted by sales.
Fine-tuning, code sandboxes and storage have their own rates, starting with LoRA fine-tuning from $0.34 per 1M tokens (SFT). All of them are in the table below.
Plan Comparison
Together AI sells services, not plan tiers. Each one's listed rate:
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Service | Price | Unit | Notes |
|---|---|---|---|
| Serverless inference | From $0.04 input to $3.00 input / $15.00 output | per 1M tokens | Rate set per model |
| gpt-oss-120B (serverless) | $0.15 input / $0.60 output | per 1M tokens | Example open model |
| Qwen3.7-Max (serverless) | $1.50 input / $4.50 output | per 1M tokens | Example larger model |
| Dedicated inference, HGX H100 | $5.49 | per GPU per hour | On demand |
| Dedicated inference, HGX B200 | $8.99 | per GPU per hour | On demand |
| GPU clusters, HGX H100 / H200 / B200 | $3.99 / $5.99 / $8.19 | per GPU-hour | On demand; reserved capacity by sales |
| LoRA fine-tuning (SFT) | From $0.34 | per 1M tokens | Minimum charge per job from $4.00 |
| Code Sandbox | $0.0446 vCPU, $0.0149 RAM | per vCPU-hour / per GiB RAM-hour | For agent code execution |
| Code Interpreter | $0.03 | per 60-minute session | |
| Shared Filesystem | $0.16 | per GiB per month | Storage |
Serverless is the right start for most teams: you pay only for tokens, and gpt-oss-120B at $0.15 input is cheap enough to test at scale. Dedicated inference makes sense only when traffic is steady enough to keep a GPU busy. Note that a GPU cluster H100 is cheaper per GPU-hour ($3.99) than a dedicated inference H100 ($5.49); the cluster is raw capacity you run yourself, while dedicated inference is a served endpoint.
Hidden Costs and Considerations
- Output tokens on large models. Kimi K3 output costs $15.00 per 1M tokens, five times its input rate. Long answers on large models add up fast.
- Idle GPUs still bill. Dedicated inference and clusters are charged per hour, used or not.
- Fine-tuning minimums. Each fine-tuning job has a minimum charge from $4.00, so many small experiments cost more than their token count suggests.
- Agent extras. Sandboxes, code sessions and storage are billed on top of tokens.
- Quoted items. Reserved capacity and GPUs not on the list need a sales quote, so they cannot be compared from the page.
Cost Estimates by Team Size
These scenarios assume monthly volumes at the listed rates.
Solo developer (serverless gpt-oss-120B).
- Input: 10 million tokens x $0.15 = $1.50.
- Output: 2 million tokens x $0.60 = $1.20.
- Total cost: $1.50 + $1.20 = $2.70 per month.
Small team of 5 (serverless Qwen3.7-Max).
- Input: 100 million tokens x $1.50 = $150.
- Output: 20 million tokens x $4.50 = $90.
- Total cost: $150 + $90 = $240 per month.
Mid-size team of 20 (one dedicated endpoint plus fine-tuning).
- Dedicated inference: 1 HGX H100 for 720 hours x $5.49 = $3,952.80.
- Fine-tuning: 50 million tokens of LoRA SFT x $0.34 = $17.
- Total cost: $3,952.80 + $17 = $3,969.80 per month.
The jump from the small-team case to the mid-size case is the move from tokens to an always-on GPU. Stay on serverless until your token bill approaches what a dedicated GPU costs per month.
How Together AI Pricing Compares
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Provider | gpt-oss-120B per 1M tokens | H100 per GPU per hour | Fine-tuning |
|---|---|---|---|
| Together AI | $0.15 input, $0.60 output | $5.49 dedicated inference; $3.99 cluster | LoRA SFT from $0.34 per 1M tokens |
| Groq | GPT OSS 120B: $0.15 input, $0.60 output | Not listed | Not listed |
| Fireworks AI | Per-model rates in its docs | $8.00 on demand | LoRA SFT $0.50 per 1M training tokens up to 16B |
| Replicate | DeepSeek R1: $3.75/1M input + $0.01/1K output tokens | $5.49 per H100 hardware hour, metered per second | Supported models only; model-specific training charges |
On gpt-oss-120B, Together AI and Groq charge the same rate. Replicate lists H100 hardware at $5.49/hour, metered per second; most private hardware deployments also bill setup and idle time. Compare deployment terms, measured utilization and the selected model's units before treating GPU hourly rates as equivalent workload costs. Its LoRA fine-tuning entry rate is below Fireworks AI's. The weak point is complexity: with separate rate cards for tokens, GPUs, fine-tuning and tooling, a mixed workload takes more work to forecast.