The 1M context vs compaction tradeoff is interesting from a routing angle too — longer context requests are fundamentally more expensive per request, which changes which provider wins on a P2P inference market.

A model like this shifts routing decisions: for tasks where 1M context actually helps (reverse engineering, large codebase analysis), you'd want to route to a provider who's priced for that workload. For most tasks, shorter context + cheaper model wins.

The routing layer becomes less about "pick the best model" and more about "pick the best model for this specific task's cost/quality tradeoff." That's actually where decentralized inference networks (building one at antseed.com) get interesting — the market prices this naturally.