OpenRouter internals: OSS models are ~75% of usage but only ~25% of spend, 80-90% traffic is agents

mobileraj · x · 2026-11-05

Takeaways from a conversation with OpenRouter's Peter Walker:

OSS models account for 75% of usage but only 25% of spend
OpenRouter runs like an RTB ad exchange: it pings providers every 5 seconds for compute, load, latency and per-token cost, then routes each request — a real-time inference marketplace
80-90% of traffic is driven by agents
Running batch jobs at night saves significant cost, analogous to setting a max CPC in ad buying
50 neocloud applications per week; many are tiny operations, some possibly proxying hyperscalers
Model fraud is real: some neoclouds claim to serve top models while actually serving lesser ones for arbitrage
The #1 reason users switch models is capability, not cost — reinforcing the shift from consumption-based to outcome-based pricing (Cursor, Fin, Sierra, etc.)

Author's read: this mirrors early programmatic advertising — fragmented supply, price variance, arbitrage and fraud, eventually consolidation. If the AI capex cycle turns, the long tail of neoclouds could face a bloodbath, with hyperscalers (cheaper capital, more purchasing power) the biggest beneficiaries.

Original post →

OpenRouter internals: OSS models are ~75% of usage but only ~25% of spend, 80-90% traffic is agents

More from Infra