A researcher at Anthropic spent $20,000 on a project to write a C compiler using its AI model Opus 4.6. After nearly 2,000 sessions and 100,000 lines of code, the result was a working C compiler that can build Linux kernel versions for multiple architectures. However, the experiment left the creator feeling uneasy due to the high costs and uncertain performance of the generated code.
The project used “agent teams,” where multiple AI instances worked in parallel on shared codebases without human intervention. This approach showed promising results but also highlighted challenges such as ensuring high-quality tests and avoiding unnecessary output.
Experts pointed out that while the project demonstrated autonomous development capabilities, it raised concerns about software deployment with unverified code. The experiment’s outcome sparked debate among developers and users, with some questioning the project’s cost-benefit ratio and others appreciating its novelty and potential benefits.
While Opus 4.6 was designed to generate high-quality code, the resulting C compiler had limitations in terms of efficiency and Rust code quality compared to expert-written code. The experiment serves as a reminder of the complexities and challenges associated with AI-powered development and the need for careful consideration of costs, risks, and performance.
Source: https://www.theregister.com/2026/02/09/claude_opus_46_compiler