# Together AI — AI Compare Ninja

> Cloud platform where developers run open AI models via API, fine-tune them and rent dedicated GPUs.

## Overview
- **Category:** Development & Code (Agent & RAG frameworks)
- **Status:** Active · checked - Active. It offers a cloud platform for running AI models (inference), fine-tuning them, and using GPU clusters. Billing is usage-based only, with no subscription tiers.
- **Paid from:** $0.03 / After consumption (Serverless Inference (Pay-as-you-go))
- **Free access:** No free plan published
- **Website:** https://www.together.ai/
- **Canonical URL:** https://www.aicompare.ninja/en/tool/together-ai

## What it does
- Run, fine-tune, and host open models on rented GPUs
- Serverless inference plus a Batch API at half the real-time rate
- Dedicated single-tenant GPUs with autoscaling for custom models
- Reserved GPU clusters (H100, H200, and others) and provisioned throughput
- Image, video, audio, transcription, embeddings, rerank, and moderation besides text
- Fine-tuning (LoRA or full), code sandbox, and shared filesystem

## In brief
Together AI is a cloud for open models: inference via API (application programming interface), fine-tuning, and rented GPUs (graphics processing units). Serverless bills by tokens; Batch costs half of real-time inference. Cached inputs are cheaper.

Dedicated Inference isolates GPUs. Clusters and Provisioned Throughput (PTU) reserve capacity. Together publishes prices in US dollars; the pricing page names no EU region. The full token and GPU list is in the pricing overview.


## Limits
Together AI is a US vendor and publishes prices in US dollars. The pricing page names no EU data region. According to the privacy policy, the vendor does not use customer data to train its own models without explicit opt-in and consent. Data may be processed outside the user's country. In Zero Data Retention mode, the vendor says inputs and outputs are not stored and are not used for model training, product improvement, or other secondary purposes.

## Pricing tiers
| Tier | Price | Period | What you get |
| --- | --- | --- | --- |
| Serverless Inference (Pay-as-you-go) | $0.03 | After consumption | Examples per 1M tokens (input/output): gpt-oss-20B USD 0.05 / 0.20; LFM2.5-8B-A1B USD 0.03 / 0.12; gpt-oss-120B USD 0.15 / 0.60; Llama 3.3 70B USD 1.04 / 1.04; Qwen3.7-Max USD 1.25 / 3.75; Kimi K3 USD 3.00 / 15.00; Cached input is cheaper, e.g. Kimi K3 at USD 0.30 instead of USD 3.00; One model is listed at USD 0.00: Ternary Bonsai 27B; Batch API at half the real-time inference price; Further modalities: image, video, audio, transcription, embeddings, rerank, moderation |
| Dedicated Inference | $5.49 / $8.99 | Per GPU and hour, on-demand | Single-tenant GPU instances with guaranteed performance; Support for custom models, autoscaling; H200, B300, GB200 NVL72 and GB300 NVL72 on request only; reserved capacity on request only |
| GPU Clusters | $3.99 / $5.99 / $8.19 | Per GPU and hour, on-demand | Reserved H100: USD 3.69 (7-30 days), 3.45 (31-90 days), 3.19 (91-180 days); Reserved H200: USD 4.99 / 4.15 / 3.99; Reserved B200: USD 7.99 / 7.79 / 6.79; 181+ days and GB200/GB300/B300 on request |
| Provisioned Throughput (PTU) | $0.05 | Per PTU and per minute | Reserved capacity in throughput units; tokens per minute delivered depends on the model; MiniMax M3: 166,667 input TPM, 833,333 cached TPM, 41,667 output TPM per PTU; Kimi K3: 16,667 / 166,667 / 3,333 TPM per PTU; GLM-5.2: 35,714 / 192,308 / 11,364 TPM per PTU |
| Fine-tuning, Sandbox and Storage | $0.48 | After consumption | Supervised fine-tuning per 1M tokens: up to 16B USD 0.48 (LoRA) / 0.54 (full); 17B-69B 1.50 / 1.65; 70-100B 2.90 / 3.20; Direct preference optimization per 1M tokens: up to 16B USD 1.20 / 1.35; 17B-69B 3.75 / 4.12; 70-100B 7.25 / 8.00; Minimum charge per fine-tuning job: USD 4.00; Code Sandbox: USD 0.0446 per vCPU-hour, USD 0.0149 per GiB RAM-hour; Code Interpreter: USD 0.03 per 60-minute session; Shared filesystem: USD 0.16 per GiB per month |

## Sources
All details regarding plans, limits, and features are based on official vendor documentation and the primary sources verified below.

- https://www.together.ai/pricing
- https://docs.together.ai/docs/billing-credits
- https://www.together.ai/privacy

---
*Source: AI Compare Ninja (2026) - https://www.aicompare.ninja/en/tool/together-ai*