If you want to spend less on AI in engineering, the levers everyone reaches for first are negotiated rates, seat counts, and model choice. All three are worth a pass. None of them move quickly. Rate negotiations run on the vendor’s calendar, seat reclamation needs a quarter of usage data plus an uncomfortable conversation, and model routing needs weeks of testing before anyone trusts it with production traffic. Quarters, not weeks.
Behavior is the fast lever. How sessions get managed, how much history gets hauled into work that doesn’t need it, how often a fresh context would have been cheaper than continuing. It accumulates every day.
Almost nobody looks at it. Engineers don’t have time to think about anything that isn’t directly affecting their productivity, and context management rarely is. In our conversations with customers about this release, engineers told us repeatedly that they had no idea what their own context patterns cost, and their managers had no way to show them even when they asked.
So two weeks before per-person context cost shipped, we turned it on for ourselves.
What we did
We switched it on across our own Claude Code dashboard. Every engineer could see their own context cost and where it sat relative to the rest of the team.
That was the whole intervention. No policy, no target, no promised review. We turned the view on and let people find it.
What happened
Within a week, gross context waste fell by roughly 300,000 tokens per session.
Context is re-sent on every turn, which is what makes that number larger than it first looks. Waste in a long session isn’t paid once. It’s paid again on every exchange for the life of the session, which is why a habit that looks like a rounding error on one request turns into a line item across a quarter. For us the reduction came to about $10,000 a month, on an engineering team at least an order of magnitude smaller than most of our customers.
Nobody was asked to change anything. People saw what their habits cost next to their peers, and they adjusted.
Why it worked
We believe this worked for two reasons.
The number was per person. Aggregate cost is somebody else’s problem by construction, and an org-wide token total tells an individual engineer nothing about what to do differently on Monday. A number with their name on it does.
The number was comparative. A raw figure is hard to act on, because there’s no scale to read it against. A position relative to the rest of the team reads instantly. People rarely argue with their own number once it sits beside everyone else’s.
We deliberately didn’t attach a consequence. We think a target would have made it worse, because the moment a number becomes a performance metric people start managing the number instead of the work. Gaming context efficiency is easy, and it would cost more in lost productivity than the context ever did. Visibility without a mandate got us the behavior change for free.
If you want to run the same test
You need three things: per-person attribution on your coding assistant spend, a comparison across the team, and the discipline to skip the policy. The third is the hard part, because a number finance can see creates pressure to attach a target to it.
Turn it on, say nothing, and look again in a week.
Per-person context cost shipped in Revenium 2.20.0, on the Adoption tab of the Claude Code dashboard. Talk to us if you want to see what your team’s context habits are costing.
