Boris Cherny, the engineer who built Claude Code at Anthropic, is challenging one of the AI industry's default ways of tracking progress. In an interview with Business Insider, Cherny argued that token burn — the volume of tokens consumed by AI systems — is a poor proxy for actual value creation, and that the industry needs to think more carefully about what it is actually measuring.
Why Token Burn Falls Short
Token burn has become a common shorthand for AI adoption. The logic is simple: more tokens processed means more usage, which suggests more value. But Cherny pushes back on that chain of reasoning. A system that burns through tokens generating incorrect code, requiring repeated corrections, may register high consumption while delivering little of worth. Volume, in other words, does not equal impact. This argument fits naturally with the broader philosophy behind Claude Code, a tool Cherny has described elsewhere as designed to handle complex, multi-step engineering work rather than simple autocomplete. As we covered previously, Claude Code's creator hasn't written a line of code himself in eight months, relying instead on the agent he built to handle the bulk of his own engineering tasks.
Key Facts
- Boris Cherny is the creator of Claude Code, Anthropic's agentic coding tool.
- He argues token burn is an unreliable measure of AI effectiveness.
- Outcome-based metrics — like tasks completed or bugs resolved — offer more signal.
- The critique applies broadly across AI products, not just coding assistants.
- Cherny has discussed similar themes around ROI and experimentation in prior interviews.
The alternative Cherny points toward is outcome-based measurement. Rather than counting tokens, teams should track whether the AI actually completed the task, reduced engineering time, or caught a bug that would otherwise have reached production. These metrics are harder to capture, but they reflect something real. The shift he is describing mirrors debates happening across the software industry about how to evaluate AI return on investment. In a related conversation covered here, Cherny addressed the tension between ROI focus and the need to keep experimenting, suggesting that pure efficiency metrics can crowd out the kind of exploratory work that produces long-term gains.
Token burn tells you how much the model is being used. It doesn't tell you whether the model is being useful.Boris Cherny, via Business Insider
A Metric Problem With Industry-Wide Implications
The critique lands at an interesting moment. AI spending is under scrutiny from investors and executives who want evidence that large model deployments are generating returns. When the primary available metric is consumption volume, companies can find themselves optimizing for activity rather than results. Cherny's framing gives engineering and product teams a concrete argument for building better instrumentation around their AI deployments. Knowing that an agent processed 10 million tokens matters less than knowing it closed 200 pull requests without human correction.
For Anthropic, this kind of thinking from a senior researcher shapes how Claude-based products get designed and evaluated internally. The concern is not abstract. Cherny has previously described managing tens of thousands of AI agents in a single day, a scale at which bad metrics can compound quickly into wasted compute and misleading conclusions. The conversation also connects to broader questions about what the next generation of software engineers will look like, a topic Cherny has addressed directly when discussing the future of the profession.
Whether the wider industry follows Cherny's lead on measurement is an open question. Token volume is easy to instrument and easy to report. Building outcome-based dashboards requires agreement on what a good outcome looks like, which varies by use case. But the argument is gaining traction, and coming from the person who built one of the most widely used AI coding agents on the market, it carries some weight. The pressure to show real results, not just high consumption numbers, is only likely to grow as AI budgets face tighter scrutiny heading into 2026.
