TL;DR
The short version
As work moves from chat to agents, token bills spike and the reflex is to cap usage and cut the bill. That pushes people back toward low-value work. The fix is a better unit: not cost per token, but cost per accepted task.
No honest per-token comparison exists across providers — tokenizers, model paths, and harnesses all differ — so only dollars per finished result compares them. Once you meter that way, the playbook is three moves: kill the tokens that spin, tune the ones that produce, and protect the ones that teach. Framing drawn from Nofar Gaspar on The AI Daily Brief.
Built on Nofar Gaspar and Nathaniel Whittemore, The AI Daily Brief (Operator's Cut), 4 Aug 2026.
The sticker price vs the operating metric
Per-token price is the sticker; cost per accepted task is what you actually run on. Because every lab tokenizes differently, models take different paths to an answer, and the harness around the model varies, a raw per-token number can't compare providers honestly. Only total cost divided by accepted results can.
The counterintuitive proof: a model that is cheaper per token can cost more per task by needing more iterations. In Databricks' own benchmark, a 1.7x-cheaper-per-token model burned 1.9x the tokens and came out more expensive per completed task.
Databricks coding-agent benchmark (verified).
Teach, produce, spin — act in that order
Sort every token into one of three jobs. Spin is machines talking to themselves — idle agents, over-frequent jobs, unread automations, immortal sessions, bloated context, unfiltered data pulls. Produce is the work that ships. Teach is the experimentation and context-building that compounds into returns.
Kill spin first, tune production second, protect teaching last — and hardest. Spin is usually well-intentioned, not error: sensible defaults left running. One idle agent burned $1,500 in two weeks at a roughly 2,600-to-1 input-to-output ratio — proof that the leak is an operations problem, not a bug hunt.
The most expensive token is the one that your best person is afraid to spend.
Nofar Gaspar, The AI Daily Brief
The reasoning-token ambush
Tokens aren't priced equally. Input is cheapest, output runs three to five times input, and a hidden reasoning layer — invisible in the answer — is billed at the high output rate. A 400-token answer can sit atop 4,000 thinking tokens.
It's the part of the restaurant bill labelled kitchen time: you never see it, but you pay a lot for it. Reasoning tokens are billed at output rates across every major provider, so a short answer can carry a reasoning bill several times its size.
Spend wisely, not sparingly
The anxious reflex — capping usage to protect the bill — trades a visible line item for invisible foregone returns, and steers teams back to low-value work. The operator move is the opposite: meter per accepted task, then defend the experimentation budget on purpose.
