Anthropic’s enterprise customers have spent two years buying tokens and a shorter time explaining the spend to a finance committee. On Oct. 28 the company shipped a set of reports meant to answer the second question.
The tools, available in the Claude administrative console, let a manager pick a date range and see what a team did with the model, what the team produced, what it cost, and where the workflow broke down. Anthropic says the reports add recommendations for improving the result, drawn from internal analytics and employee surveys, and that the underlying data will not be used to train models by default. A second feature lets an organization pool budget across teams so spending is drawn from one account rather than a card per department.
The same day, Anthropic released a version of Claude aimed at lawyers, which handles document search, comparison and drafting inside the tools a firm already uses. Legal work has been among the earliest beachheads for frontier models, and a crowded set of startups already sells into firms that bill by the hour.
What the reports actually sell is a number. Anthropic told investors this week that net revenue retention reached 900%, which means customers that were paying in the third quarter of 2025 spent about nine times as much a year later. The company also recorded an operating profit in the second quarter, a result that gives it something no rival has yet put in front of public shareholders.
Nine times is not a figure a chief financial officer absorbs quietly. A tool that was a pilot in one department becomes a line item in four, and the person signing the purchase order starts asking what the line item replaced. Anthropic’s reports are built for that conversation, and the timing is not accidental: the company filed to go public this month, and every enterprise contract it can defend is worth more in the document than one it cannot.
The clearest outside evidence of what those contracts contain comes from Jellyfish, a Boston company whose engineering-management platform pulls activity data out of tools such as Git and Jira. Its October report covers more than 1,300 companies, 276,000 developers and 100 million pull requests. Tools built on Anthropic’s models, Claude Code first among them, accounted for nearly a third of the code changes submitted for review at the median company.
The spending curve in that data is steeper than the adoption curve. Jellyfish puts Claude Code spending at a median of $63 per active developer per week, or $252 on a four-week month, with the 90th percentile at $1,666 a month. Spending on the tools has grown about tenfold since January and now equals roughly 2% of total engineering cost.
“Even if you’re saying 2% is still small, that 10x growth is growth that most CFOs didn’t plan for,” Nicholas Arcolano, Jellyfish’s chief scientist, said in an interview. Once a budget line grows at that rate, the argument shifts from whether to buy to how to control it.
Control is the harder problem, because token pricing is a meter rather than a license. A seat costs the same whether the person uses it or not; a model bills for every retry, every long context window and every agent loop that runs to completion and produces nothing. Managers who cannot see the meter running discover the cost at the end of a quarter, which is exactly the moment Anthropic’s reports arrive.
Whether the reports measure anything that matters is a separate question. Counting pull requests, sessions and tokens is easy because the platform sees all of it. Counting defects avoided, review time saved or a release that shipped without an incident is hard, and much of that evidence lives in systems Anthropic does not own. The employee surveys fill part of the gap with self-report, the softest form of proof available.
There is a structural conflict in a vendor grading the return on its own product. Anthropic built the dashboard, chooses the metrics, writes the recommendations and sells the tokens the dashboard accounts for. Buyers who have watched cloud cost-management tools mature over the past decade will recognize the pattern: the first version shows spend, the second shows waste, and the third is where the vendor’s incentive and the customer’s diverge.
The pooled budget has its own politics. Centralizing the spend moves the decision from individual engineering managers to a platform or finance group, which is efficient for the buyer and useful for the seller. A single pooled account makes renewal a conversation with one person, and it makes expansion a matter of raising a ceiling rather than winning a new champion.
The competitive pressure behind all of this is straightforward. Anthropic leads among enterprise buyers, and the stickiness comes less from benchmark scores than from the accumulated work of wiring a model into internal systems. A reporting layer adds to that work by making Anthropic’s console the place where usage is defined, which raises the cost of moving to another lab even if the models converge.
The lawyer tool points in the same direction. A vertical product bundles the model with the workflow, the security posture and the audit trail a regulated buyer needs, and it converts a general-purpose API into something a firm can defend to its own risk committee.
The open question is what happens when the meter becomes visible. If customers use the reports to cut seats and cap agents, the 900% retention number that Anthropic is showing investors becomes the peak of a curve rather than its slope. If they use it to justify expansion, the reports will look in hindsight like the cheapest product the company ever built.
Anthropic has given buyers a dashboard and kept the pricing schedule. The next quarter of enterprise spending, and the pricing of an offering that depends on it, will show which of those two facts managers acted on.
