Jensen Huang says a developer earning $500,000 should be burning at least $250,000 a year on AI tokens, and that he'd be "deeply alarmed" if they weren't. Meta's engineers ran 60.2 trillion tokens in a single month, roughly $900 million at list API prices, and the company ranked them on an internal leaderboard before quietly killing it. Salesforce went the other way, capping Claude Code at $250 per developer per month, then scrapping the cap to "reduce friction."
Different reactions, same mistake. All of them are guessing.
Here's the thing nobody wants to say out loud: spending isn't the problem. If a developer ships in a day what used to take a week, there is no token bill too high to justify it. The problem is that right now, no one can tell the difference between a session that moved the company forward and a session that burned the same tokens and produced nothing.
You're either being told to spend more, or being told to spend less. Neither camp has the data to know which is right.
That's the gap we built AI Coding Insights to close. It connects what you spend on Claude Code and Cursor to what that spending actually produced, so you can finally answer the questions a billing total can't: which teams, sessions, and models are delivering real value, which patterns are worth spreading, and whether to spend more or less. The rest of this post is how. First, why the bill alone will never get you there.
The bill tells you the total. It tells you nothing else.
If you've rolled out Claude Code across a team, you already know the shape of this. The Anthropic billing tab gives you one number: total spend. No per-session view. No per-developer view. No sense of whether the spend turned into merged code or into an agent running in circles for four hours.
So the questions that actually matter go unanswered:
- How much are we spending on coding agents, and on what?
- What is each dollar actually delivering, and should we be spending more or less?
- Which teams and individuals are getting real value, and what are they doing differently?
- When we change the model or the setup, does value-per-dollar improve?
There's a deeper reason these questions stay unanswered. Coding agents are moving on three fronts at once: the agents themselves, the harness and configuration around them, and the underlying models. Change any one and both your spend and your output shift, but with telemetry scattered across each agent's own billing tab, you can't tell which change did what. Did merges speed up because you moved to Fable, or because of the new permission setup? Today most teams are flying blind. A unified observability layer across your coding agents is what unlocks those comparisons: this agent against that one, this harness change against the last, this model against the cheaper one, all measured the same way.
Cost on its own is a trap
Here's where most cost dashboards stop, and why they mislead.
Say one model takes five times as many turns to reach a good result, but costs ten times less per token. Is it cheaper or more expensive? On the bill, it looks cheap. But if those extra turns mean branches sit open twice as long before they merge, it's quietly the most expensive option you have.
You cannot answer that with cost data alone. Cost is only meaningful next to what it produced and how long it took. When spending goes up because you're shipping more, that's healthy. When it goes up because a model is inefficient, that's waste and it looks identical on the invoice.
That's the gap. And closing it is the whole point of what we're launching.
Introducing AI Coding Insights
Today we're launching AI Coding Insights, starting with full support for Claude Code and Cursor. More coding agents are coming soon.
It gives engineering leaders the same visibility into their coding agents that they already have into their production services. It rests on four connected pillars:
| Pillar | The question it answers | What you see |
|---|---|---|
| Cost | Where are the tokens actually going? | Spend per session, per developer, per repo, per model. Sort by cost to find the sessions that ran in circles. Cache savings broken out. |
| Adoption | Is this taking hold, or stalling? | Active developers and reach, session depth, model usage by team, and whether the rollout is spreading or concentrated in one corner. |
| Productivity | Are tokens spent efficiently? | Session turns, tool-call breakdown, and conversation depth as a team-wide trend, plus cycle time from branch to merge by model. |
| Quality | How is AI-generated code quality evolving? | How many attempts it takes to reach a useful result, code-review comments per PR, changes requested before merge, and merge rate. |
Let the plugin capture what the bill can’t
Claude Code already speaks OpenTelemetry. So does Dash0.
Install the open-source Dash0 Plugin for Claude Code from our website github.com/dash0hq/dash0-agent-pluginand session telemetry starts flowing as standard OTel traces. The plugin captures 26 event types across 11 categories, including subagents, MCP tool calls, and permission requests, so you see what each session actually did, not just a token count at the end.
Within minutes you can open AI Coding Insights, sort every session by cost or error count, and read the full conversation exactly as Claude Code ran it: every prompt, every tool call, every subagent, every denied permission. Connect GitHub, and cycle time lights up next to spend.
Because it's all OpenTelemetry underneath, none of it is a walled garden. Query any metric, build custom dashboards, set alerts, and correlate agent activity with the downstream services it touches, all in the same platform you already use for production.
Token maxing, but you can actually see it
We're not here to tell you to spend less. If the data says a team is shipping faster because they're leaning into Fable, spend more. If it says half your Fable budget is going to planning tasks Sonnet handles just as well, you finally have the receipt.
The point was never to cap the bill. The point is to stop guessing, so that when you push the spend up, it feels good, because you can see exactly what you're getting back.
Your software factory is already running on AI. Now you can see inside it.