Nicholas Carlini has Claude Opus 4.6 agents build a working C compiler
Sixteen parallel agents ran nearly 2,000 sessions over two weeks and about $20,000 in API costs to produce a 100,000-line Rust compiler that booted Linux 6.9 on three architectures.
- Models & capabilities
- Benchmarks & progress
- Minor
Nicholas Carlini, a researcher on Anthropic’s Safeguards team, directed 16 parallel instances of Claude Opus 4.6 to build a C compiler from scratch, working largely without direct human supervision. The brief was demanding: a GCC-compatible compiler, written in Rust with no external dependencies beyond the Rust standard library, capable of building a bootable Linux 6.9 kernel on x86, ARM and RISC-V.
Over nearly two weeks and almost 2,000 Claude Code sessions, at a cost Anthropic put at just under $20,000, the agents produced a roughly 100,000-line compiler. It compiled and booted the Linux kernel across all three target architectures, built other real-world projects including QEMU, FFmpeg, SQLite, PostgreSQL and Redis, passed 99% of the GCC torture-test suite, and could compile and run Doom.
Anthropic’s write-up was explicit about the result’s limits. The compiler lacked a 16-bit x86 compiler needed for real-mode booting, so it relied on GCC for that step; it depended on external assemblers and linkers rather than implementing its own; and its generated code was less efficient than GCC running with all optimisations disabled. The exercise was published as an engineering case study of “agent teams” — many model instances collaborating on one codebase — rather than as a benchmark result, and it followed Opus 4.6’s release earlier in the month as one of several public demonstrations of the model’s capacity for long, loosely supervised coding work.