Together AI launches Together Link for coding-agent model routing

Together AI launched Together Link in beta, giving developers a way to connect the coding-agent harnesses and desktop apps they already use to models on its inference platform. Together Link adds automatic model routing and usage-cost tracking without requiring teams to switch tools.

The setup documentation lists support for Claude Code, Claude Desktop and Cowork, Codex CLI, ChatGPT Desktop, OpenCode 2, and Pi Code 0.80.8 or newer. Installation uses a single shell command and requires macOS or Linux, a Together API key, and a supported tool that is already installed.

Together AI says Link points each tool directly to its hosted gateway instead of running a local proxy or daemon. Terminal agents receive temporary configuration for each launch, while desktop apps use separate Link profiles.

The company’s materials differ on how often automatic routing occurs. Its launch post and product page say Link classifies the first task and selects a model once for the session. The documentation, however, says the gateway classifies and routes each request. Users can bypass the router by pinning a model for a session. The documentation lists Kimi K3, GLM 5.3, GLM 5.3 Flash, and DeepSeek V4.1 Flash among the Together-hosted choices.

Optional routing to Anthropic applies only to Claude Code and Claude Desktop sessions, according to the documentation. Codex, OpenCode, Pi Code, and ChatGPT Desktop remain on Together-hosted models. So do Claude sessions without an Anthropic API key.

Together AI says Link itself is free, while inference on its platform is billed to the user’s Together API key at serverless rates. Requests sent to Claude Opus are billed separately through the user’s Anthropic account. Link prints a cost summary when a session ends, and its usage command reports gateway-tracked spending across sessions.

The company claims the router can cut coding-agent model spending by more than 50%. Its product FAQ gives a range of 50% to 80% compared with running every session on Opus 5.5. The company materials provide no reproducible methodology or independent validation for that range. Together AI also describes the available open models as frontier-quality, but the materials provide no independent benchmark for that broader performance claim. The product page says Together Link is open source under the MIT license.