Anthropic Publishes Economic Research Into How Developers Use Claude Code at Scale

Anthropic has published economic research examining how developers use Claude Code — the company’s AI-assisted coding environment — drawing on detailed usage logs and telemetry data gathered at scale across the product’s user base.

The research asks three specific questions: who is using Claude Code and for what kinds of tasks; how the value of those tasks is shifting over time; and how much a developer’s domain expertise shapes whether a coding session succeeds or fails.

What Claude Code Actually Tracks

Claude Code automatically writes detailed local usage logs covering token counts, models used, sessions, and projects across Anthropic’s API, Pro, and Max plans. Organisations running Team or Enterprise accounts can access an Analytics section inside the Claude Console — Anthropic’s developer dashboard — where they can monitor adoption and output across their entire engineering team.

The metrics available are specific. Organisation-level analytics include lines of code accepted, suggestion accept rate, daily active users and sessions, and a daily breakdown of accepted code lines over time. At the individual level, the Console records each team member’s total lines of code accepted from Claude Code suggestions in a given month.

Organisations can go further still. Claude Code supports telemetry export via OpenTelemetry — the open-source observability standard — allowing companies to stream anonymised metrics, logs, and optional traces into their own monitoring systems for time-series analysis of usage, costs, and tool activity.

How Developers Are Actually Using It

Community analysis of months of Claude Code usage logs — shared by independent developers and not officially verified by Anthropic — offers a granular picture of day-to-day behaviour. Around half of all Claude Code calls reportedly use the lower-cost Haiku model, likely because developers lean on it for large-file work where cost efficiency matters.

The Edit tool accounts for about 35% of all tool interactions. Read comes second at around 22%, with TodoWrite — Claude Code’s built-in task-tracking tool — at roughly 18%.

These figures are indicative rather than official. But they’re consistent with the kinds of questions Anthropic’s research is trying to answer.

The Productivity Measurement Problem

Anthropic has framed Claude Code explicitly as a tool for improving what it calls “engineering velocity” — the speed and output of software development teams. The analytics framework is built to measure that impact in concrete terms.

When organisations install the Claude GitHub App and enable GitHub analytics, they can link Claude Code usage directly to software delivery metrics: pull requests opened per user, contribution rates, and code accepted per session. That connection between AI tool usage and actual software output is what makes the research genuinely economic in character — it treats developer time, AI assistance, and domain knowledge as inputs with measurable outputs.

A growing ecosystem of third-party tools has emerged alongside the official Console. Tools like `ccusage`, Claude-Code-Usage-Monitor, and a VS Code extension from Analytic Projects all help developers interpret their own logs, track personal token consumption, and estimate costs. The `claude-usage` project transforms local logs into charts and cost breakdowns. Their existence signals real developer demand for understanding the value — and the bill — of AI-assisted coding.

Concerns Around Monitoring

Not everyone is comfortable with the direction of travel. Some developers and commentators have raised concerns that detailed usage tracking — chiefly when linked to individual contribution metrics and pull request counts — risks becoming a tool for workplace surveillance rather than genuine productivity insight.

There’s also a methodological question. Lines of code accepted and suggestion accept rates are clean numbers, but critics caution they don’t necessarily capture code quality, architectural thinking, or the kind of slow, careful work that prevents bugs downstream. Leaning too hard on quantitative proxies for productivity is a risk any organisation using these dashboards should keep in mind.

Adam Jermyn, a researcher at Anthropic, has previously written about how AI tools change the nature of technical work — though the specific economic research referenced here represents a distinct strand of the company’s effort to understand its own products in use.

What This Means for Kent Residents

For software developers and digital teams across Kent — whether working in logistics, fintech, or public services — this research matters in practical terms: it shows that organisations using Claude Code via Anthropic’s Team or Enterprise plans already have detailed analytics available to measure whether AI-assisted coding is actually speeding up their work. Public sector digital teams, including those at Kent County Council or NHS Kent and Medway ICB, may find the productivity and cost-tracking tools relevant when evaluating whether AI coding assistance offers genuine value for money — though any deployment would need to clear data and security policies first. For individual developers in Kent on personal subscriptions, free third-party tools like `ccusage` can already show exactly how much each session is costing and which models are doing the heavy lifting.