Together AI

Together AI is a cloud platform for running, training and fine-tuning open-source and open-weight AI models. Developers reach it through an API and a Python SDK and CLI, and it targets teams that want models such as Qwen, Kimi, GLM, DeepSeek, MiniMax and gpt-oss without operating their own GPU fleet.

What it offers

Serverless Inference is pay-per-token access to a catalog of chat, vision, image, audio, video, transcription, embedding, rerank and moderation models, with no provisioning step. Prices are set per model and per million tokens, and many chat models carry a separate, lower cached-input rate. Provisioned Throughput reserves fixed capacity in throughput units for teams with steady traffic. Dedicated Inference runs single-tenant GPU instances (for example NVIDIA HGX H100 and B200) billed per GPU-hour, with support for custom models and autoscaling. GPU Clusters rent on-demand, preemptible or reserved GPU capacity by the hour. The platform also sells code sandboxes and a code interpreter, a shared filesystem for clusters, post-training on B200 hardware, and LoRA and full fine-tuning billed per token processed.

API compatibility

Together exposes an OpenAI-compatible API, and in November 2026 it added support for the Anthropic Messages API and the OpenAI Responses API for selected GLM models, so an app built on either can switch by changing its API key, base URL and model ID.

Model lifecycle

Together publishes a model lifecycle policy. Close upgrades of the same model lineage can be redirected after a short notice, while releases with materially different behavior or a price increase above 10% are treated as new models with no automatic redirect. Serverless deprecations usually get two to three weeks of notice by email and are listed with removal dates and recommended replacements on its deprecations page. Per-model serverless prices can change, so teams should check the live pricing page rather than relying on older documentation.

Who it is for

Developers and AI product teams that want one account and one bill across many open models, plus a path from serverless calls to dedicated endpoints, fine-tuning or reserved GPUs as usage grows.

Pricing type: paid, usage-based. Alternatives include Fireworks AI, DeepInfra, Baseten, OpenRouter (a router across providers) and first-party model APIs such as Alibaba Cloud Model Studio.

Official site: https://www.together.ai