Together AI pricing: serverless tokens, GPUs and fine-tuning
Checked Nov 6, 2026 on Together AI’s official pages (sources below). Prices change often: confirm with the provider before you buy.
Together AI hosts open-source and partner models behind a serverless API, rents GPU clusters, and runs fine-tuning jobs. Serverless inference bills per 1M tokens (with cheaper cached input on many models); GPUs bill per GPU hour.
Figures below are quoted from together.ai/pricing, checked Nov 6, 2026. The page lists dozens of models; we quote a sample.
Serverless inference: sample chat models (per 1M tokens)
- gpt-oss-120B
- Input$0.15
- Cached input—
- Output$0.60
- GLM-5.3
- Input$1.40
- Cached input$0.26
- Output$4.40
- GLM-5.3-Flash
- Input$0.15
- Cached input$0.03
- Output$0.50
- Kimi K3
- Input$3.00
- Cached input$0.30
- Output$15.00
- DeepSeek V4.1 Flash
- Input$0.30
- Cached input$0.006
- Output$1.20
- Llama 3 8B Instruct Lite
- Input$0.14
- Cached input—
- Output$0.14
Audio: Whisper Large v3 $0.0015 per audio minute; Cartesia Sonic-3 TTS $65.00 per 1M characters. Images from $0.0017 per megapixel (Juggernaut Lightning Flux).
GPU clusters (per GPU per hour)
- NVIDIA HGX H100
- Preemptible$1.99
- On demand$3.99
- Reserved 3–30 days$3.69
- Reserved 91–180 days$3.19
- NVIDIA HGX H200
- Preemptible$2.99
- On demand$5.99
- Reserved 3–30 days$4.99
- Reserved 91–180 days$3.99
- NVIDIA HGX B200
- Preemptible$4.09
- On demand$8.19
- Reserved 3–30 days$7.99
- Reserved 91–180 days$6.79
- NVIDIA HGX B300
- Preemptible$4.99
- On demand$9.99
- Reserved 3–30 daysContact us
- Reserved 91–180 daysContact us
Dedicated model inference is listed separately: HGX H100 $5.49 and HGX B200 $8.99 per GPU hour on demand.
Fine-tuning (LoRA, per 1M training tokens)
- Llama 3.1 8B
- Supervised fine-tuning$0.34
- DPO$0.84
- Minimum charge$4.00
- Llama 3.3 70B
- Supervised fine-tuning$2.03
- DPO$5.08
- Minimum charge$4.00
- gpt-oss-20B
- Supervised fine-tuning$0.40
- DPO$1.00
- Minimum charge$4.00
- gpt-oss-120B
- Supervised fine-tuning$2.50
- DPO$6.25
- Minimum charge$6.00
Billed on training tokens (dataset size × epochs) plus evaluation tokens; each job has the minimum charge shown.
Serverless, dedicated or clusters?
Serverless bills per token and suits most apps. Dedicated inference rents single-tenant GPUs for a model with guaranteed performance. GPU clusters are raw capacity by the GPU hour, cheapest when preemptible or reserved for longer terms. Together also sells provisioned throughput units (PTUs) and code sandboxes ($0.03 per 60-minute Code Interpreter session). Compare Groq pricing and Replicate pricing.
Together AI pricing FAQ
How much does Together AI cost?
It depends on the product. On together.ai/pricing (Nov 6, 2026), serverless gpt-oss-120B was $0.15 per 1M input tokens and $0.60 per 1M output; an on-demand H100 GPU cluster was $3.99 per GPU hour; LoRA fine-tuning of Llama 3.1 8B was $0.34 per 1M tokens.
Does Together AI charge less for cached input?
Yes, on many models. For example GLM-5.3 lists $1.40 input, $0.26 cached input and $4.40 output per 1M tokens.
Does Together AI rent GPUs by the hour?
Yes. GPU clusters are listed per GPU per hour, with preemptible, on-demand and reserved (3 to 180+ days) prices.
Sources (official pages, read Nov 6, 2026)
Spot something out of date? Email hello@youraitools.co and we’ll fix it.