Uber's CTO confirmed it publicly: they burned through their entire 2026 AI coding budget by April. Four months in, done. Claude Code adoption had jumped from 32% to 84% across roughly 5,000 engineers, and monthly API costs landed somewhere between $500 and $2,000 per engineer.
Microsoft responded by revoking Claude Code licenses across its Experiences and Devices division, redirecting engineers to GitHub Copilot CLI by June 30. GitHub is pivoting Copilot to usage-based billing. And according to TechCrunch's June 5 piece, the FinOps Foundation is fielding what they're calling "existential crises" from companies running 3x over their token budgets.
This is what it looks like when token spend finally catches up with adoption.
Token spend is invisible until the bill arrives
The Uber situation is extreme, but it's not a fluke. It's the predictable outcome when teams move fast on AI tooling without any instrumentation on what each run actually costs. And it isn't a Fortune 500 problem. Any team running AI coding agents — Claude Code, Codex, Copilot, Cursor — in CI or locally is accumulating costs with zero visibility into what each run actually costs. The problem isn't AI adoption. The problem is that agentic workflow cost is invisible until the bill arrives.
Your Anthropic dashboard shows you aggregate monthly spend. Your Cursor invoice shows seats. What neither tells you is that issue #47 burned $12.40 across four agent runs including a failed retry that re-ran the entire context. Or that your deploy-staging workflow costs 6x more per run than run-tests because of how the agent is prompted. Or that one contributor is responsible for 40% of your Claude Code cost because they're running agents on every draft PR.
That's not a complaint about Anthropic's billing interface — it's just not what that tool is for. But without per-run attribution, you're flying blind on the metric that's about to determine whether your team gets to keep using these tools.
What teams actually need isn't a simpler invoice
It's run-level observability. You need to know:
Which workflows are your biggest spenders? A failed agent run that retries three times before timing out costs the same in tokens as a run that succeeds — but only one of those runs delivered value. When you can see cost-per-run alongside success/failure status, the inefficient workflows become obvious immediately.
Which repos are expensive? If you have ten repos integrated with agent workflows and two of them are responsible for 70% of your token spend, you need to know that before it's a budget emergency.
Which contributors are driving spend? This isn't about punishing people for using AI — it's about understanding whether expensive usage correlates with better output. If your highest-spending contributor is also shipping the highest-quality code, that's a very different conversation than if the spend is concentrated in abandoned experiments. (The Uber COO's admission that they can't draw a line between token consumption and shipped features is exactly this problem at scale.)
Is the money working? Spending more on AI coding agents should improve something measurable — cycle time, review quality, defect rate. If you can't connect agent spend to outcomes, you're operating on faith.
AgentMeter
AgentMeter exists to fill this gap.
The core is a GitHub Action that captures token usage from every agent run — input tokens, output tokens, cache reads, model, cost — and associates it with the specific workflow, repo, and triggering event. Setup is three lines of YAML:
- uses: agentmeter/action@v1
with:
api-key: ${{ secrets.AGENTMETER_API_KEY }}
run-name: ${{ github.workflow }}After every run, the action posts a cost summary directly on the PR: what it cost, how many tokens, how it compared to recent runs on that workflow. No context-switching to a separate dashboard to understand what just happened.
The dashboard surfaces everything aggregated: trend charts so you can see week-over-week cost trajectory, per-workflow breakdowns so you know which workflows to optimize first, per-contributor costs so you can have informed conversations about usage, and quality ratings so you can start connecting spend to outcomes.
It's free for individual repos. Teams are $19/month. The goal is to make AI agent cost tracking something you set up in five minutes and forget about until you actually need the data — not another tool that requires its own operational overhead.
What's next
The piece coming next is local session tracking. Right now AgentMeter captures what happens in CI — every agent run that goes through GitHub Actions. But a significant chunk of AI coding agent spend happens locally: Claude Code sessions, Cursor agent runs, Codex sessions before they ever hit CI. Surfacing that in the same dashboard means teams can see total agent spend across both environments, not just the CI slice.
That's in progress. In the meantime, CI visibility alone is enough to make teams significantly more deliberate about their agentic workflow cost — and to avoid the kind of April surprise that Uber's engineers had to explain to their CTO.
If you're running AI coding agents in GitHub Actions and you're not tracking per-run costs yet, agentmeter.app is where to start.