The coding agent adoption number that stopped meaning anything

A founder I know opened her last board deck with one number: 88 percent of engineers actively using a coding agent. The room nodded. It felt like progress. Then her finance lead pulled up the first metered invoice and the same 88 percent looked like something else entirely. Not a victory lap. A spend forecast nobody had set a ceiling on.

That shift, from adoption-as-trophy to adoption-as-cost, finished landing this week.

The myth that high adoption means you are winning

For most of the past year, the coding agent story leadership teams told themselves was simple. Get the tools into more hands, watch the adoption curve climb, report the percentage upward, feel good. Adoption was the proxy for value because adoption was free. Flat seats, all-you-can-eat sessions, one predictable line on the budget. The more people used it, the better the deal looked.

So the instinct made sense. If a tool is cheap and fixed, wide usage is pure upside.

Why adoption looked like value under flat pricing

It sounds right because it was right, for exactly as long as the pricing held. When a seat costs the same whether an engineer runs two tasks a week or two hundred, then yes, more usage is more value extracted from a fixed cost. Adoption percentage was a clean, honest signal under flat pricing. Nobody was paying for the marginal task, so every additional active user genuinely improved the return.

The trouble is that the pricing model the metric depended on quietly disappeared. And a metric built on an assumption that no longer holds does not announce its own retirement. It just keeps showing a green arrow while the ground underneath it changes.

The cost bands that turned adoption into exposure

Look at what actually shipped in the last few days. On June 18 a cost reference from morphllm restated the spend bands that now define this category, drawing on Anthropic’s own enterprise figures: roughly $13 per developer on an active day, $150 to $250 per developer per month on average, and power users between $500 and $2,000 a month. Ninety percent of users stay under $30 on any given active day. Read that distribution again. It is not flat. It is steep, skewed, and entirely individual. The bill now depends on which engineers use the agent and how hard, not on how many seats the company bought.

As morphllm put it, summarizing the enterprise math: “Claude Code averages $150 to $250 per developer per month,” with heavy automation users reaching “$500 to $2,000 per month.”

Then look at where the whole category is heading. On June 16 Microsoft made Copilot Cowork generally available, and the design tells the story. It is metered in credits at one cent each. It is disabled by default. It ships with spending limits at the tenant, group, and user level, plus alerts when a budget hits 80 percent. Microsoft is not selling usage. It is selling the ability to govern usage. When the vendor builds the cost ceiling into the product before anyone asks for it, the vendor already knows that wide adoption without governance is a problem its customers are about to have.

And on the same June 18, the other end of the same trend showed up. Google’s Gemini CLI stopped serving free, AI Pro, AI Ultra, and individual Code Assist users. Access narrowed to enterprise and Cloud license holders, with everyone else pushed to a closed replacement. Free, ungoverned automation is being withdrawn precisely where nobody was watching the meter. Scripts and pipelines calling the old command simply broke.

Three vendors, one direction. The cost of adoption stopped being zero, and it stopped being hidden.

Verified output per dollar of agent spend

Here is the better mental model, and it is calmer than the panic version. Adoption is not the win. Adoption is the input. The win is verified output per dollar of agent spend, measured on the engineers actually generating it.

That means two numbers belong on the same slide, side by side. On one line, the spend per active engineer per month, which the meter now hands over for free. On the other, the verified, merged work that engineer produced with the agent. Not lines of code, not pull requests opened, not session count. Work that shipped and held up. The ratio between those two lines is the only thing that says whether usage is value or just usage.

The question stopped being how many engineers use the agent. It became which of them turn agent spend into shipped work, and whether anyone can see the difference.

This reframe also dissolves a fear I keep hearing in the same week as the adoption brag. Leaders worry that capping spend will throttle their best people. The skewed distribution says otherwise. Most engineers cost very little. A handful drive the bill. A per-engineer ceiling does not punish the team. It surfaces the heavy users so a leader can ask the one question worth asking: are they the highest-output people, or just the highest-spend ones? Sometimes those are the same engineers. Often they are not, and nobody could tell before the meter existed.

Three moves for your next budget review

For anyone with a board update or a budget review this quarter, three small moves do most of the work.

Replace the adoption percentage on the slide with cost per active engineer next to verified merged output. If only one of those lines is buildable this month, build the cost line, because the meter already wrote it. Then set a per-engineer monthly ceiling and say the number out loud, so people stop quietly self-rationing in the dark, skipping the exploratory runs that the meter taxes first. And name one person who owns the call when the next pricing change lands, because there will be one.

None of this is a fire drill. The tools still work, the engineers still ship, and the metering shift is genuinely good news in one respect: for the first time, the cost side of the ratio is honest. The number that got worse is the old slide. The number that got better is the one a finance team can finally compute. The companies that come out of this calm are not the ones with the highest adoption. They are the ones who stopped mistaking it for the scoreboard.