Quick answer
The local whoburnedmore CLI command is npx whoburnedmore --local. Decline its optional publish offer to keep the report local. The standard command signs in and submits daily aggregates. Neither mode submits prompts, source code, or filenames.
What this guide answers
The reader wants better metrics than tokenmaxxing leaderboards alone.
Best for
Engineering leaders evaluating AI coding programs.
Key facts, checked November 8, 2026
- METR's July 2025 randomized trial found experienced open-source developers took 19% longer with AI tools, while believing they had been about 20% faster. source
- METR's February 2026 follow-up measured an 18% speedup for returning developers and 4% for new ones. METR warned that selection effects make the data unreliable and that the true speedup is likely larger. source
- The 2025 DORA report found AI adoption now has a positive link to delivery throughput but still a negative link to delivery stability. 90% of respondents use AI at work. source
- DORA's tokenmaxxing guidance (June 2, 2026) recommends pairing usage with existing outcome metrics such as cost per accepted change and code rework rate. source
Burn is not productivity
Token usage is easy to count and easy to game. It should start questions, not end them.
- What shipped?
- What improved?
- What failed?
- What should change?
Use paired metrics
Every usage metric needs an outcome metric beside it. This keeps teams from rewarding waste and helps defend genuinely valuable high usage.
- Tokens plus merged PRs.
- Cost plus cycle time.
- Agent runs plus review load.
Measure review burden
AI-generated code can move effort from writing to reviewing. ROI should include whether reviewers are spending more time catching defects.
- Diff size.
- Review comments.
- Rework rate.
- Escaped bugs.
Keep qualitative notes
Some AI value is hard to see in dashboards. Capture examples where AI helped understand unfamiliar code, unblock a migration, or prevent toil.
- Incident story.
- Migration story.
- Learning story.
Why self-reported speedups are not enough
The METR results are the strongest warning about perceived ROI. In the 2025 study developers expected a 24% speedup and still felt faster afterwards, but they were measurably slower. The 2026 follow-up found speedups, but METR itself says the data is unreliable. Between 30 and 50% of surveyed developers withheld tasks they did not want to do without AI. The takeaway for ROI work: surveys show sentiment, and you need system data to show throughput and quality.
- Use surveys for satisfaction and friction
- Use system data (PRs, cycle time, rework) for throughput
- Never use either alone
Paired metrics with an official data source
Each pair below can be pulled from a vendor dashboard or API rather than built by hand. Claude Code's Team and Enterprise analytics show lines accepted, suggestion accept rate and sessions, with GitHub contribution metrics in public beta. Its OpenTelemetry export emits claude_code.cost.usage, claude_code.token.usage, claude_code.pull_request.count and claude_code.commit.count. Copilot usage metrics have a dashboard and REST API, generally available since February 27, 2026. Cursor team analytics shows the AI share of committed code.
- Cost per merged PR: claude_code.cost.usage divided by merged PRs
- Rework rate: share of AI-touched lines rewritten within a window you pick, such as three weeks
- Stability: change failure rate before and after rollout, per DORA
- Adoption: weekly active users from each vendor dashboard
Where the burn number comes from
Vendor dashboards each cover one tool. For one input figure across tools, npx whoburnedmore reads local logs from 22 coding agents and reports daily tokens and estimated cost by tool and model. Divide that by your merged-PR count for the same window to get a cross-tool cost per change.
Common questions
What is the best metric for AI coding ROI?
No single metric works. DORA recommends pairing usage with outcome metrics such as cost per accepted change, rework rate and delivery stability, measured at team level.
Does AI actually make developers faster?
The evidence is mixed. METR's 2025 trial found a 19% slowdown for experienced developers. Its 2026 follow-up found speedups but called the data unreliable. Measure your own team's outcomes.
Can I export Claude Code usage to my metrics stack?
Yes. Claude Code supports OpenTelemetry (enabled with CLAUDE_CODE_ENABLE_TELEMETRY=1), with metrics for cost, tokens, sessions, commits and pull requests.