How Anthropic runs large-scale code migrations with Claude Code
02:00 · July 16, 2026 · Claude Blog
Summary
Anthropic engineers have shown that large-scale code migrations, once multi-year efforts requiring dedicated teams and parallel codebases, can now be completed by individuals in days or weeks with current frontier models. In recent projects, developers at the company used Claude Fable 5 and Claude Opus 4.8 to port ten packages ranging from tens to hundreds of thousands of lines. The work relied on dynamic workflows that combined model classes for planning, implementation, review, and verification rather than treating the models as simple code generators.
Two cases illustrate the change in scope. Jarred Sumner migrated the Bun runtime from Zig to Rust, producing roughly one million lines of code in under two weeks while achieving full passage of the existing test suite before merge. Nineteen post-merge regressions were identified and resolved, after which the Rust version shipped inside Claude Code. The effort consumed 5.9 billion uncached input tokens and 690 million output tokens, corresponding to approximately $165,000 at prevailing API rates. Mike Krieger completed a separate migration of an internal Python tool to 165,000 lines of TypeScript over a single weekend, incorporating hundreds of subagents, eight phase gates, three adversarial review cycles, and an exhaustive output-diff parity check against the original implementation.
The economic threshold for such projects has shifted because the dominant cost is no longer sustained engineering time but token consumption and iteration overhead. A migration can now be justified by concrete operational gains, such as reducing binary compilation from roughly eight minutes per platform to two seconds, achieving a sixfold improvement in startup latency, and retiring a separate deployment pipeline. The same principle applies when chronic issues like memory-safety maintenance or a single performance bottleneck become the deciding factor.
The central lesson drawn from these migrations is that effort should focus on refining the generation and verification loop itself rather than on manual correction of individual code fragments. Fable 5 and Opus 4.8 proved effective at decomposing work across subagents, maintaining multiple solution paths, and enforcing consistency checks, which allowed the human operator to intervene primarily at the level of process design.
Why it matters
This article provides Product Teams and Builders with concrete, actionable insights into using advanced LLMs for large-scale code migrations. It includes specific models, token costs, and strategic frameworks that Dutch AI practitioners can adopt to modernize legacy systems efficiently.
