# Together AI API pricing and models

Together AI publishes per-token prices for 31 text models, plus 1 free variant, as of October 20, 2026. The lowest output rate is $0.14 per million tokens, on Llama 3 8B Instruct Lite, and the highest-scoring model it serves is Kimi K3, at 79 on our 100-point score.

Together AI runs a serverless API for open-weight models and rents GPU clusters. This page covers the serverless per-token prices only.

Source: [Together AI pricing page (serverless table)](https://www.together.ai/pricing), read 2026-10-20. Prices in US dollars per million tokens.

## Models we score

| Model | Score | Input | Output | Cached input | Context | vs maker |
|---|---|---|---|---|---|---|
| Kimi K3 (`moonshotai/Kimi-K3`) | 79 | $3.00 | $15.00 | $0.30 | 1.05M | Same |
| Qwen3.7 Max (`Qwen/Qwen3.7-Max`) | 76 | $2.50 | $7.50 | $0.50 | — | — |
| DeepSeek V4.1 Flash (`deepseek-ai/DeepSeek-V4.1-Flash`) | 73 | $0.30 | $1.20 | $0.006 | 1M | Same |
| DeepSeek V4.1 Flash 1017 (`deepseek-ai/DeepSeek-V4.1-Flash-1017`) | 73 | $0.30 | $1.20 | $0.006 | 1M | Same |
| Qwen3.7 Plus (`Qwen/Qwen3.7-Plus`) | 71 | $0.32 | $1.28 | — | 1M | — |
| GLM-5.3 Flash (`zai-org/GLM-5.3-Flash`) | 69 | $0.15 | $0.50 | $0.03 | 1.05M | — |
| DeepSeek V4 Pro 0813 (`deepseek-ai/DeepSeek-V4-Pro-0813`) | 69 | $1.32 | $3.96 | $0.13 | 1.05M | Same |
| Beam 1.0 501B A23B (`reflection/Beam-1.0-501B-A23B`) | 68 | $0.40 | $1.20 | $0.04 | 1.05M | — |
| Qwen3.6 Plus (`Qwen/Qwen3.6-Plus`) | 68 | $0.50 | $3.00 | — | 1M | — |
| GLM-5.3 (`zai-org/GLM-5.3`) | 68 | $1.40 | $4.40 | $0.26 | 1.05M | — |
| Inkling (`thinkingmachines/Inkling`) | 67 | $1.00 | $4.05 | $0.17 | 524K | -13% |
| Minimax M3 (`MiniMaxAI/MiniMax-M3`) | 65 | $0.30 | $1.20 | $0.06 | 524K | Same |
| GLM-5.2 (`zai-org/GLM-5.2`) | 65 | $1.40 | $4.40 | $0.26 | 1.05M | Same |
| DeepSeek V4 Flash 0731 (`deepseek-ai/DeepSeek-V4-Flash-0731`) | 62 | $0.14 | $0.28 | $0.03 | — | Same |
| Llama 3.3 70B (`meta-llama/Llama-3.3-70B-Instruct-Turbo`) | 61 | $1.04 | $1.04 | — | — | — |
| Muse Glimmer 30B (`meta-models/Muse-Glimmer-30B`) | 61 | $0.35 | $1.50 | $0.04 | 131K | — |
| MiniMax M2.7 (`MiniMaxAI/MiniMax-M2.7`) | 60 | $0.30 | $1.20 | $0.06 | — | Same |
| Gemma 4 31B (`google/gemma-4-31B-it`) | 56 | $0.39 | $0.97 | — | — | — |
| GPT-OSS 120B (`openai/gpt-oss-120b`) | 55 | $0.15 | $0.60 | — | 131K | — |
| Ternary Bonsai 27B (`Prism-ML/Ternary-Bonsai-27B`) | 0 | $0 | $0 | — | 262K | — |
| Tev1 4B Experimental (`together/Tev1-4B-experimental`) | 0 | $0.042 | $0 | — | 33K | — |

## Other models

| Model | Input | Output | Cached input | Context |
|---|---|---|---|---|
| Llama 3 8B Instruct Lite (`meta-llama/Meta-Llama-3-8B-Instruct-Lite`) | $0.14 | $0.14 | — | — |
| Rnj-1 Instruct (`essentialai/rnj-1-instruct`) | $0.15 | $0.15 | — | — |
| Qwen3.5 9B (`Qwen/Qwen3.5-9B`) | $0.17 | $0.25 | — | 262K |
| Qwen2.5 7B Instruct Turbo (`Qwen/Qwen2.5-7B-Instruct-Turbo`) | $0.30 | $0.30 | — | — |
| Qwen3.8 Flash (`Qwen/Qwen3.8-Flash`) | $0.15 | $0.47 | — | 1M |
| NVIDIA Nemotron 3.5 Super VL (`nvidia/NVIDIA-Nemotron-3.5-Super-VL`) | $0.15 | $0.50 | $0.02 | 262K |
| Qwen3 235B A22B Instruct 2507 FP8 Throughput (`Qwen/Qwen3-235B-A22B-Instruct-2507-tput`) | $0.20 | $0.60 | — | — |
| Cogito v2.1 671B (`deepcogito/cogito-v2-1-671b`) | $1.25 | $1.25 | — | — |
| Qwen3.5-397B-A17B (`Qwen/Qwen3.5-397B-A17B`) | $0.60 | $3.60 | $0.35 | — |
| Muse Spark Open 816B A42B (`meta-models/Muse-Spark-Open-816B-A42B`) | $1.25 | $4.25 | $0.15 | 1.05M |
| Qwen3.8-2.4T-A95B (`Qwen/Qwen3.8-2.4T-A95B`) | $2.00 | $6.00 | $0.50 | — |

## What we measured (2026-09-01)

All 900 requests succeeded, and Together’s GLM-5.2 p95 was second only to OpenRouter. Its invoice charged 2,928,000 GLM-5.2 input tokens against 1,667,041 reported by its own API responses, 76% more, while the same method reconciled to the token on DeepSeek Flash. We record this as observed and unexplained; the findings section has the figures, and we have invited a correction.

| Model run | Invoiced /1M blended | Median latency | p95 latency | Succeeded |
|---|---|---|---|---|
| DeepSeek V4 Flash | $0.111 | 2.55 s | 10.09 s | 100.0% |
| GLM-5.2 | $1.667 | 2.53 s | 5.66 s | 100.0% |

Method and invoices: https://benchlm.ai/openrouter-alternatives

## Hosts that sell the same models

- [OpenRouter](https://benchlm.ai/hosts/openrouter): 18 shared, 12 cheaper per output token (DeepSeek V4.1 Flash 1017: $0.12 vs $1.20)
- [DeepInfra](https://benchlm.ai/hosts/deepinfra): 13 shared, 11 cheaper per output token (DeepSeek V4.1 Flash 1017: $0.29 vs $1.20)
- [Novita AI](https://benchlm.ai/hosts/novita): 13 shared, 4 cheaper per output token (Llama 3.3 70B: $0.40 vs $1.04)
- [Nebius Token Factory](https://benchlm.ai/hosts/nebius): 10 shared, 1 cheaper per output token (Beam 1.0 501B-A23B: $1.00 vs $1.20)
- [Baseten](https://benchlm.ai/hosts/baseten): 8 shared, 2 cheaper per output token (GPT-OSS 120B: $0.50 vs $0.60)
- [Fireworks AI](https://benchlm.ai/hosts/fireworks): 6 shared, 0 cheaper per output token

## FAQ

### How much does the Together AI API cost?

Together AI bills per million tokens, with separate input and output rates for each model. On October 20, 2026 its output rates ran from $0.14 on Llama 3 8B Instruct Lite to $15.00 across 31 paid models. Input is usually cheaper than output, so a long prompt with a short reply costs less than the output rate suggests.

### Which models can you run on Together AI?

32 text models are listed, and 20 of them match a model we score. The highest-scoring is Kimi K3 at 79. The table on this page has every model with its input, output and cached-input rate, the context window, and the date we read the price.

### Is Together AI fast?

In our paid September 1, 2026 run, Together AI returned a median of 2.55 seconds on DeepSeek V4 Flash and 2.53 seconds on GLM-5.2, over 450 requests per model. That is one workload on one date, so check your own prompts before you rely on it.

### What are the alternatives to Together AI?

OpenRouter lists 18 of the same models and charges less output per token on 12 of them. DeepInfra and Novita AI also overlap. Price is only part of the choice: speed, rate limits and which model snapshot a host serves differ too.

Canonical page: https://benchlm.ai/hosts/together