A key per provider, edited one at a time#
Model Router offers a slot for each of the 13 providers it can actually route to, and each slot is edited on its own, so adding a fifth key never touches the other four. A card says plainly when a provider is accepted but not yet routable. Keys can be added after onboarding, and a hybrid mode falls back to managed capacity when your own key cannot serve a request.
An org model policy#
Block a model for the whole organization and it leaves every routing tier, whichever key would have served it. The provider list comes from the same registry the router reads, so the settings page and the router can never disagree about what is available.
ROI and usage, over the same window#
A new ROI tab compares routed cost with what the same work would have cost sent direct, over the same window as the rest of the dashboard. Usage-over-time charts show router-scoped spend with full token columns, and `router stats` in the CLI now leads with your most recent session.
Improvements#
- Model RouterChoose the auto top-up amount and the balance floor that triggers it.
- Model RouterOpenRouter, Azure AI Foundry and Fireworks AI as bring-your-own-key providers.
- Model RouterOrg members, not only admins, can view the router dashboard.
- Model RouterEach conversation gets its own prompt cache.
- CLIThe router daemon runs on Windows.
Fixes#
- Model RouterAuto top-up charges the saved card as intended.
- Model RouterSavings are no longer floored per row, so the two dashboard pages agree.
- Model RouterThe daily cache-hit rate can no longer exceed 100%.
- Model RouterThe Sessions and Members tabs honour the range picker.
- CLI`router on` always works for organizations on their own keys.