Anthropic Publishes Cost Ledger for Bun's Million-Line AI Migration

Bun's team ported over a million lines of code from Zig to Rust in under two weeks for $165,000 in AI agent costs, per Anthropic's engineering blog. The bill: 5.9 billion input tokens against 690 million output, an 8.5-to-1 ratio. Reading code, not writing it, is the real cost of an agentic migration.

The port shipped as a single 1,448-file pull request, with the full existing test suite passing before merge. Nineteen regressions surfaced afterward, all fixed, about one for every 53,000 lines touched. Anthropic doesn't say how long the fixes took or what the regressions cost Bun's users in the meantime.

A second migration in the same post tells the sharper story. It ported 165,000 lines from Python to TypeScript over a weekend, using hundreds of agents behind eight phase gates and three adversarial review rounds. Total cost: 27 million tokens, for roughly a sixth of Bun's line count and a fraction of Bun's cost per line. Phase gates and adversarial review, not model quality, explain that gap. Gates stop agents from re-reading the same files across independent attempts. A review pass built to find failure catches what a single pass waves through.

Anyone budgeting an agentic rewrite should treat cost as reading volume times number of passes. Put a real adversarial review stage before merge, not after.