Silicon Valley’s AI token boom hides costly code rewrites

The new status symbol in Silicon Valley is not shipping software faster, but burning more AI tokens. Engineers using Claude Code, Cursor and Codex are generating more accepted code, yet managers may be mistaking that burst of volume for durable productivity when later rewrites strip away much of the gain.

Waydev CEO Alex Circei says the illusion is built into the dashboard. Managers may see AI code acceptance rates of 80% to 90%, he said, but once later revisions are counted, the real-world acceptance rate can fall to between 10% and 30%. Waydev, which works with 50 customers employing more than 10,000 software engineers, has reworked its platform over the last six months to track AI-agent metadata and code quality and cost signals, a sign that the industry is still searching for metrics that capture what happens after the first draft lands.

The economics are getting harder to ignore. Jellyfish said it analyzed 12,000 developers across 200 companies in the first quarter of 2026 and found the median developer used about 51 million AI coding tokens a month, while the 90th percentile used about 380 million. Jellyfish estimated monthly token spend at about $52.38 at the median and about $691.14 at the 90th percentile. At the extremes, the gap in output was far smaller than the gap in cost: developers in the bottom 20% of token spend averaged 11 merged pull requests over the quarter and spent about $3 on tokens, while the top 20% averaged 23 merged pull requests and spent $1,822.

That is the managerial blind spot now spreading through software teams. High token consumption can look like speed, but it can also mean more code to review, more code to correct and more technical debt to clean up later. The result is a productivity myth that rewards visible volume while hiding the labor of rewriting work that was never quite finished.

The caution flags are not new. McKinsey & Company found in 2023 that software developers could complete coding tasks up to twice as fast with generative AI, but only when teams also changed their processes and upskilled workers. An April 2025 academic paper reached a similar conclusion, finding GenAI coding assistants were strongest on routine tasks and weaker on complex, domain-specific work because they lacked context and support for custom design rules.

The culture around AI coding has started to mirror that obsession with volume. Reports in April described internal leaderboards at Meta with titles such as Token Legend, Model Connoisseur and Cache Wizard. Other reports said one OpenAI engineer processed 210 billion tokens in a week, while an Anthropic Claude Code user ran up a bill of more than $150,000 in a month. The lesson is not that AI cannot help engineers write code. It is that token burn alone says little about whether the code will survive the next review, the next rewrite or the next cost review.