Ordalin
Submit a tool

Reviewed tool profile

Together AI

Cloud platform to run, fine-tune and serve open-source AI models through an OpenAI-compatible API.

Sources checked

Visit website ↗
Pricing
Pay-as-you-go usage pricing that varies by product and model. Serverless language models are billed per million input and output tokens (for example $0.15 in and $0.60 out for gpt-oss-120B), media models per image, video or audio minute, dedicated endpoints and GPU clusters per GPU-hour, and fine-tuning per million training tokens with a per-job minimum. Reserved GPU capacity and the largest GPU systems go through sales. No ongoing free plan is described, although a few models are listed at $0.00. Official source 1 ↗
Available as
A developer platform used through its OpenAI-compatible API, SDK integrations and a CLI, plus a web Playground and Together Chat for trying models. GPU clusters support Slurm. Official source 1 ↗ · Official source 2 ↗
Primary group
Coding & Development

What it does

Together AI is a cloud platform for running, fine-tuning and serving open-source AI models. Developers call a large library of chat, code, image, video and voice models through an OpenAI-compatible API, with serverless, batch, reserved-throughput and dedicated deployment options. The same platform rents GPU clusters for training, offers fine-tuning on your own data, code sandboxes and managed storage, and Together Link lets you use its open models inside the coding harness you already work in.

Official source 1 ↗ · Official source 2 ↗

Features

  • Serverless inference on open models. Run open-source models on demand through an OpenAI-compatible API with no infrastructure to manage and no long-term commitment. The catalogue spans chat, vision, image, audio, video, transcription, embeddings, rerank and moderation models. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗
  • Batch inference. Process large workloads asynchronously at lower cost, up to 30 billion tokens per model, with any serverless model or a private deployment. Official source 1 ↗
  • Provisioned Throughput. Reserve committed inference capacity in throughput units (PTUs) with token-based pricing and a 99% uptime SLA, keeping drop-in API compatibility. Official source 1 ↗ · Official source 2 ↗
  • Dedicated model and container inference. Deploy models on single-tenant GPUs with guaranteed performance, custom-model support and autoscaling. Dedicated Container Inference targets video, audio and image models. Official source 1 ↗ · Official source 2 ↗
  • GPU clusters. Rent NVIDIA GPU capacity from self-serve instant clusters up to thousands of GPUs, including H100 and B200 clusters with attached storage for training or large batch jobs, on preemptible, on-demand or reserved terms. Official source 1 ↗ · Official source 2 ↗
  • Fine-tuning. Fine-tune open-source models on your own data with supervised fine-tuning or Direct Preference Optimization, using LoRA or full fine-tuning, then deploy the result for inference. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗
  • Together Link. Run Together's open models inside the coding harness you already use, with one install command, automatic routing and per-session cost receipts. It is marked beta in the docs. Official source 1 ↗ · Official source 2 ↗
  • Code sandboxes and managed storage. Secure code sandboxes provide development environments for AI apps and agents, a Code Interpreter API runs LLM-generated code, and managed object storage and parallel filesystems have zero egress fees. Official source 1 ↗ · Official source 2 ↗
  • Data-use controls. Together AI says it does not train its models on your data without explicit opt-in, and account privacy settings can stop it retaining training data, prompts or model responses. The docs also cover zero data retention and single sign-on. Official source 1 ↗ · Official source 2 ↗

Pricing & free limits

Serverless inference (language models)

From $0.15 input / $0.60 output per 1M tokens (gpt-oss-120B); varies by modelPay as you go, per 1M tokens

Each model has its own input and output token price, often with a cheaper cached-input rate and a separate Batch API price. Examples: gpt-oss-120B $0.15 / $0.60, MiniMax M3 $0.30 / $1.20 (cached input $0.06), Llama 3.3 70B $1.04 / $1.04. Official source 1 ↗

Serverless inference (image, video, audio)

e.g. $0.053 per image, $0.115 per video, $0.0015 per audio minutePay as you go, per image, video, audio minute or 1M characters

Media models are priced per output unit: GPT Image 2 at $0.053 per image, ByteDance Seedance 2.5 at $0.115 per video, Whisper Large v3 transcription at $0.0015 per audio minute. Text-to-speech is priced per million characters, for example Kokoro-82M at $4.00. Official source 1 ↗

Provisioned Throughput

Priced per throughput unit (PTU); estimate on the pricing pageReserved capacity, continuous provisioning

You reserve a number of PTUs, each a fixed slice of capacity whose tokens per minute depend on the model. The pricing page includes a cost estimator; estimates assume 24/7 provisioning. Comes with a 99% uptime SLA. Official source 1 ↗ · Official source 2 ↗

Dedicated Inference

$5.49 per GPU-hour (H100); $8.99 (B200)On-demand per GPU per hour; reserved through sales

Single-tenant GPU endpoints. H200, B300 and GB200/GB300 NVL72 hardware and all reserved capacity are quoted by sales. Official source 1 ↗

GPU Clusters

$3.99 per GPU-hour on demand (H100); $1.99 preemptiblePer GPU per hour; lower rates for reserved terms

H100 runs $1.99 preemptible, $3.99 on demand, and $3.69 to $3.19 for 3- to 180-day reservations; longer reservations and NVL72 systems are by quote. Other on-demand rates: H200 $5.99, B200 $8.19, B300 $9.99. Official source 1 ↗

Fine-tuning

From $0.34 per 1M training tokens (small models); varies by model and methodPer 1M tokens processed, with a per-job minimum

Supervised fine-tuning and DPO are priced per model, from $0.34 (SFT) and $0.84 (DPO) per million tokens on small Qwen models up to $40 and $100 on GLM-5.2. Each job has a minimum charge, from $4.00. Official source 1 ↗

Sandbox and storage

$0.0446 per vCPU-hour; $0.03 per Code Interpreter session; $0.16 per GiB-month storageUsage-based

Code Sandbox VMs cost $0.0446 per vCPU and $0.0149 per GiB of RAM per hour. Code Interpreter sessions of 60 minutes cost $0.03. Shared filesystem storage is $0.16 per GiB per month. Official source 1 ↗

Billing details

Fine-tuning cost is based on the tokens in the training dataset times the number of epochs, plus any evaluation-set tokens. Official source 1 ↗

Listed image and video prices are for the lowest resolution and duration settings; actual prices can be higher. Official source 1 ↗

Free access & limits

Free access and its limits were not established by the checked official sources.

Best for & limitations

Building an app or agent on open-source models through an OpenAI-compatible API without running inference infrastructure. Official source 1 ↗ · Official source 2 ↗

Customising an open model with your own data and serving the fine-tuned model in production. Official source 1 ↗

Renting H100 or B200 GPU clusters for model training or large batch jobs. Official source 1 ↗

Using cheaper open models inside an existing coding harness through Together Link. Official source 1 ↗

  • Newest GPU systems and reservations by quote. H200, B300 and GB200/GB300 NVL72 dedicated endpoints, NVL72 clusters and long reservations have no public price and require contacting sales. Official source 1 ↗
  • Together Link is in beta. Together Link is labelled beta in the documentation. Official source 1 ↗
  • Media prices are minimums. Listed image and video model prices apply to the lowest resolution and duration settings; actual costs vary. Official source 1 ↗
Together AI product screenshot
Website previewFixed desktop viewport · 1440 × 900

Similar tools

Published tools that share its categories.

OctriTurns one OpenAPI spec into AI-drafted API docs, SDKs in ten languages, an MCP server and production monitoring.
Coding & DevelopmentFree planChecked Oct 5
Visit