AI EngineerVideoSource Linked
Notion's Token Town — Sarah Sachs, Notion
Agent economics can regress even when token prices look stable. Route by task, preserve model optionality, and move deterministic work out of LLM calls before scaling usage.
AI Engineer · Jul 23, 2026
Source Summary
Notion describes recurring upgrade traps: unchanged per-token pricing paired with **3× output tokens**, or a replacement model costing **40% more** while its predecessor is retired within **four months**. Automatic upgrades can therefore raise costs without proportional product value.
Practical Implication
Route each agent task to the cheapest model that meets your own quality, latency, and tool-use requirements. Preserve **model optionality**, evaluate open-weight alternatives, and use ordinary code for deterministic transformations, CLI calls, and SQL queries.
Agent-Ready Context
Notion describes recurring upgrade traps: unchanged per-token pricing paired with **3× output tokens**, or a replacement model costing **40% more** while its predecessor is retired within **four months**. Automatic upgrades can therefore raise costs without proportional product value. Route each agent task to the cheapest model that meets your own quality, latency, and tool-use requirements. Preserve **model optionality**, evaluate open-weight alternatives, and use ordinary code for deterministic transformations, CLI calls, and SQL queries. Open-weight models may cover more tasks over time, but the talk does not claim they lead every workload today. Savings depend on internal evals, interoperability work, governance, and knowing which jobs genuinely require frontier reasoning.
Context Map
infracoding#model-selection#gateways#open-modelsUncertainty
Open-weight models may cover more tasks over time, but the talk does not claim they lead every workload today. Savings depend on internal evals, interoperability work, governance, and knowing which jobs genuinely require frontier reasoning.