What changed
Together AI launched Together Link in beta on October 5. The free, MIT-licensed tool connects Claude Code, Claude Desktop, ChatGPT Desktop, Codex CLI, OpenCode and Pi to models served by Together AI while preserving the user's existing harness, settings and normal provider setup. Terminal integrations use temporary/per-session configuration; desktop integrations use reversible profiles. The default Auto Router chooses a model once per session, preserving prompt caching, and each proxied session can report token and dollar totals.
Why it matters
Coding-agent products increasingly bundle two distinct layers: the harness that manages tools/context/workflow and the model that reasons. Together Link makes that boundary operationally visible by letting developers keep a familiar harness while changing the model/provider underneath it. That can make cost and model experiments easier without retraining a team on another agent interface. The economics need caution: Together says open-model use can cut agent spend by more than half or 50–80%, but that is a vendor claim and cost per token is not the same as cost per successfully completed task.
The harness stays; the model layer moves
Together Link supports Claude Code, Claude Desktop, ChatGPT Desktop, Codex CLI, OpenCode and Pi. Rather than asking teams to adopt another coding agent, it adapts those existing interfaces to Together-served models.
Configuration is deliberately reversible
For terminal agents, Together says configuration is applied only for the launched session. Desktop integrations use separate profiles that can be turned off, reducing the switching cost of testing another model provider.
Routing happens once per session
Auto Router examines the initial session task and chooses a model for that session rather than rerouting every turn. Together says this preserves prompt caching. With an optional Anthropic API key it can also route harder work to Opus 5.5; without one it stays on Together-served models.
Cost claims need task-level evidence
Together markets savings of more than 50%, and its product page says 50–80% versus running every session on Opus 5.5. Those figures are vendor-produced. Independent harness research reinforces that the harness can materially affect results and cost, which is precisely why token-price comparisons alone cannot establish cost per solved task.