Editorial illustration for Track AI cost per active developer and merged pull request, not tokens
AI Coding Tools: Measure Cost Per Developer, Not Tokens
Track AI cost per active developer and merged pull request, not tokens
Uber burned through its entire 2026 AI coding budget by April, according to reports, after developers leaned harder into Claude Code. GitHub Copilot has since swapped its flat-rate plan for metered credits. Gartner puts a date on where this is headed: by 2028, token costs for AI coding tools will outpace average developer salaries.
The math behind that is simple enough. A coding agent doesn't just answer a prompt once, it plans, writes code, tests it, finds the bugs, and loops back to fix them. That cycle can chew through 10 to 50 times the tokens of a single chat response.
So the agents doing the most useful work end up generating the biggest bills, and finance teams are noticing.
The instinct is to clamp down, usage caps, seat limits, approval gates before anyone can spin up an agent. HPE's latest piece on engineering economics argues that's the wrong fix, because it measures the wrong thing. A per-token meter can't tell the difference between a runaway retry loop and the five-minute agent session that just shipped a fix to production. Capping spend blind doesn't cut waste, it just as easily cuts the work that matters.
A better measure is cost per merged pull request: total AI-assisted development cost divided by the number of changes that pass review and are merged. Track rework and defects alongside it so cheaper changes do not conceal lower quality.
Why this matters
For teams running agentic coding tools, the token bill is a symptom, not the disease. The real question is whether that spend turns into merged code, and most dashboards aren't built to answer it. Cost per active developer and cost per merged pull request give engineering leaders a way to separate a handful of power users from a sea of idle licenses, and to tie dollars spent to work actually shipped.
That matters because the instinct to ration agent calls when a bill spikes is probably backwards: if an agent's iterative planning and error-correction cycle is what makes it useful, cutting tokens just makes it worse at the job without fixing the underlying cost structure. Running a full development cycle on owned capacity, as HPE suggests, is a useful stress test. If per-developer and per-PR costs don't improve on infrastructure you control, the problem isn't pricing, it's how the workflow is built.
For founders and engineering leads evaluating AI coding tools, that's the metric to demand from vendors, not token counts dressed up as efficiency.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives The Letter Company Releases Letterpress-1, a Coding Agent for Growth Engineering
From Archives Goldman Sachs cuts traders and bankers off from advanced AI coding tools
Common Questions Answered
Why did Uber exhaust its entire 2026 AI coding budget by April?
Uber's developers leaned heavily into Claude Code, which increased token consumption significantly. Coding agents perform multiple iterations including planning, writing, testing, and bug fixing, causing token costs to accumulate rapidly beyond initial budget projections.
What metric does the article recommend instead of tracking token costs?
The article recommends tracking cost per merged pull request, which is calculated as total AI-assisted development cost divided by the number of changes that pass review and are merged. This metric should be tracked alongside rework and defects to ensure cheaper changes do not conceal lower quality outcomes.
When does Gartner predict token costs for AI coding tools will surpass average developer salaries?
According to Gartner, by 2028, token costs for AI coding tools will outpace average developer salaries. This projection reflects the exponential growth in token consumption as coding agents perform iterative tasks like planning, writing, testing, and bug fixing.
How did GitHub Copilot change its pricing model in response to rising AI coding costs?
GitHub Copilot swapped its flat-rate plan for metered credits, shifting from a subscription model to usage-based pricing. This change reflects the industry's move toward more granular cost tracking as token expenses become increasingly unpredictable.
What two key metrics should engineering leaders use to evaluate AI coding tool spending?
Engineering leaders should track cost per active developer and cost per merged pull request to understand their AI-assisted development spending. These metrics help separate power users from idle licenses and directly tie dollars spent to work actually shipped, revealing whether high token bills translate into meaningful code contributions.
Further Reading
- Report: Claude Code Exceeds GitHub Copilot for Merging Pull Requests - DevOps.com
- Software factories: AI agents write code, people ship it - R&D World
- AI Coding ROI: Complete Cost & Measurement Guide - nOps
- Healthy code needs 22 to 35% fewer tokens for agentic coding - CodeScene
- Models and pricing for GitHub Copilot - GitHub Documentation
