What surprised me here isn’t that Claude Code can emit telemetry. It’s that the piece makes a pretty strong case that you can get from “we think spend is rising” to “show me which model, which user, and which kind of call did it” with almost no infrastructure. That’s useful. But I’m also a little wary of how clean it sounds, because a lot of observability demos look effortless right up until you try them against real teams, real auth, and real billing boundaries.
The most interesting part is not Grafana. It’s the shape of the data Anthropic chose to expose. Cost, tokens, session counts, edit decisions, and a few work classifications are exactly the kind of things that make a manager say, “oh, so that’s where the money went.” You can feel the article leaning toward a fairly practical thesis: don’t start by instrumenting everything; start by answering the boring questions people actually ask after the invoice lands. That feels right.
I do think the privacy posture is doing a lot of work here. The defaults are intentionally restrained: no prompt text unless you opt in, no bash commands, no raw file contents. That’s the part I’d trust most, honestly. If Claude Code can give you spend attribution without turning into a surveillance firehose, that’s a decent tradeoff. Still, “the docs say the email only goes to the endpoint you configure” is the sort of line I’d verify before I treated it as a blanket promise. Not because I assume bad faith, just because telemetry systems have a habit of growing edges over time.
The dashboard queries themselves are the more concrete win. “Who spends,” “which model,” and “what kind of work” are not fancy panels, but they’re the right ones. If I were trying this in anger, I’d probably start with the model split and the subagent/main split, because those are the places where intuition is usually worst. People tend to blame “AI usage” in the abstract; the labels force a more honest conversation. Maybe the expensive model is justified. Maybe the subagents are where the bill leaks. Maybe not. At least you’d know.
I’m less sure about how universally reliable the cost numbers will feel in practice. The article does mention that the figures are approximations and that official billing lives elsewhere, which is the right caveat, but that also means this is really consumption telemetry, not finance-grade accounting. That distinction matters. If someone tries to reconcile this directly against an invoice, they may be disappointed. If they use it to see trends and outliers, it sounds genuinely useful.
The other thing I’d test immediately is the label story. It’s nice that team.id can become a useful dimension with one environment variable, but the moment you move from a single laptop to a real org, identity gets messy fast. API keys, OAuth, managed settings, repo overrides, redaction rules — all of that becomes the actual product, not the Grafana panel. That’s where these setups usually either become a standard operating practice or quietly die in a month.
What I like most is that this article treats observability as a conversation starter, not a vanity dashboard. That’s the right mental model. Spend visibility in LLM tooling is valuable precisely because it turns vague anxiety into something debatable. Once you can see the shape of usage, you can argue about model choice, caching, agents, and policy instead of just shrugging at the bill.
Reference: Claude Code Already Emits OpenTelemetry: Who Spends What, on Which Model, in Grafana
