Together Link was released by Together AI through its official blog on October 5, 2026. It is a connection layer that plugs the coding agent tools developers already use into open models hosted on Together AI, unchanged. The pitch is blunt: same agent, same workflow, and a bill cut by over 50%. Setup takes one command, existing settings and logins stay as they were, and switching back to the native closed models also takes one command.
Only the model is swapped — the tool is not
Coding agents are now everyday equipment for engineering teams, yet everything from a one-line fix to a full-repo rewrite often runs on the same most expensive closed model by default. Together Link works with Claude Code, Claude Desktop, Codex in the ChatGPT app and the CLI, OpenCode and Pi, which covers the mainstream ways these agents are used today. It is not trying to build yet another coding assistant. It pulls out the model-supply layer: the interface, shortcuts and team habits stay put, and only the inference requests behind them are rerouted to open models.
Auto routing: flagships for hard problems, Flash for small jobs
What actually decides the savings is the Auto router. It reads the first task of each session to judge difficulty: quick fixes go to fast, cheap models, and only hard problems get frontier capability. Routing happens once per session, so prompt caching keeps working. If you bring your own Anthropic key, it routes between Opus 5.5 and GLM 5.3; without one, it routes between GLM 5.3 and GLM 5.3 Flash. After each session, a tracker shows what you spent next to what the same session would have cost on Opus 5.5. Billing runs on your own Together API key, pay-as-you-go or credit packs, with no separate contract.
It only works because open models caught up first
Together AI is explicit about the timing in its blog post: open-weight models such as Kimi K3 and GLM 5.3 can now handle many of the hardest coding tasks, while small models like GLM 5.3 Flash and DeepSeek V4.1 Flash cover everyday work at a fraction of the per-token price. Engineering organizations spend anywhere from tens of thousands to millions of dollars a month on closed models; once the price gap gets that wide, the routing layer becomes a business of its own. By Together AI's figures, as of September 30, 2026 it served the largest share of OpenRouter tokens for DeepSeek V4.1 Flash at about 40.8%, GLM 5.3 Flash at about 28.2% and Kimi K3 at about 23.1% — the supply side is already at scale.
The boundary deserves stating: the "over 50%" figure is the vendor's claim, and real savings depend on a team's task mix. If every session is a repo-scale rewrite, routing has little room to maneuver, and Together Link itself leaves the hardest tasks to frontier models. What it really changes is the default: instead of flagship models for everything, the first question becomes whether this task deserves a flagship at all.
