Hassan ships Together Link, a CLI that auto-routes open models
Six weeks of work, routing per session, and no benchmark numbers yet.
Hassan released Together Link in beta, a CLI that runs open models inside coding harnesses. He lists Claude Code, Codex and opencode as supported, along with others, plus auto routing that picks a model based on the task and built in spend and usage tracking. He says he and his team worked on it for the last six weeks.
What the routing actually does
The one hard detail about the router came from a reply. Connex asked how accurate the auto routing is at choosing the right model per task, and Hassan answered that routing happens per session rather than per request, that the team has found it to work well, and that benchmarks will likely come later.
Per session matters more than it sounds. A router that decides once at the start of a session is picking a model for a whole stretch of work, not re-evaluating when the task shifts from reading a repository to writing a patch to explaining a failure. That makes it cheaper and simpler to reason about, and it also means a bad first call follows you for the rest of the session. Nothing in the launch post says how a session boundary is defined or whether you can override the choice.
What is not here
There is no accuracy figure, no cost comparison, no latency number and no named benchmark attached to the release. The claim that the routing works well is the builder's own assessment of his own product, stated in a reply, with the benchmarks explicitly deferred to the future. That is worth separating from the rest of the announcement, which is a straightforward list of what the tool supports.
Spend and usage tracking is the part of the pitch that needs no benchmark to be useful. If you are already running open models through a coding harness and splitting usage across several of them, a built in meter is a thing you can verify yourself on day one.
For people building on it, treat the auto routing as the unproven half of the product until Hassan publishes numbers. Routing claims are exactly the kind of thing that has gone both ways in public this month, and a six week old beta with no eval behind it is a tool to try on a side project, not something to put in front of a paying workload. The spend tracking you can trust immediately because you can read it against your own bill. The model selection you cannot, yet, and the person who shipped it said as much when asked.
right now we do routing per session and we've found it to work quite well!
How accurate is the auto routing feature in choosing the right models to use per task?