Update routing

Refresh the model roster in the routing table that the route skill reads: $AI_ROUTING_TABLE, default ~/.agents/ai-routing.md. If it's missing, stop and say so; the starter is ai-routing.example.md next to the route skill. This is judgment work, not a find-replace. The comparisons in step 2 are where most of the value is.

Steps

1. Get current data: a fresh intelligence-vs-cost chart (or an equivalent benchmark) and a screenshot of the current model picker for each provider the table routes to. If what the user shares is cropped, missing an axis, or otherwise incomplete, ask for the rest rather than guessing at it.
2. Compare within each provider only. The table rations flat-rate subscription seats, so a chart's dollar-cost axis is pay-per-use API pricing. That's valid for comparing tiers or generations within one provider, never for ranking across providers: "this model is cheaper than that provider's model" isn't a reason to prefer it when both seats are already paid for. For each provider, note:
   • Genuine dominance: a newer tier or generation matches or beats an older one on intelligence at the same or lower cost. It replaces the old first choice.
   • A real trade-off: cheaper but also less capable. Don't force a swap. List both and say which axis moved.
   • Models from a provider the table doesn't list as a seat: mention them if relevant, but never route to them.
3. Search the web for current messages-per-5h figures per model. Date and cite every number you add. If a figure isn't published for a new model yet (common right after a launch), say so: mark it inherited from the slot the previous generation held, or as an estimate. Never present a guess as measured.
4. Update only the parts of the table that carry the roster:
   • The main phase table (bare model names, no reasoning-tier suffix).
   • The rate-limit reference card (throughput numbers, sourced and dated).
   • An efficient-settings section, if the table has one (which tier or generation is worth spending on, and why). Leave worked examples alone unless a model name changed. They reference the main table's choices by name, so tier-only changes don't touch them.
5. Bump every "Last checked" date you touched to today.
6. Run /route once afterward on a real task to check the output still reads right end to end.

What not to do

• Don't overwrite a benchmark citation with a number for a newer model unless that model was measured on the same benchmark. A different benchmark measuring a different thing isn't a substitute.
• Don't fold a cross-provider cost comparison into the reason for a routing change. It's the wrong axis for a flat-rate-seat problem.
• Don't silently drop the prior generation from the rate-limit card while the provider still offers it. Note it as the prior generation instead of deleting the row.