Summary
Boris Cherny, the creator of Claude Code, sits down with Y Combinator for a wide-ranging conversation about Anthropic’s newly shipped Opus 5 model and what it means for developers building with AI. The interview opens with Opus 5’s performance on ARC-AGI 3 — hitting 30%, up from low single digits — and Cherny’s observation that the model can run autonomously for days or weeks without external scaffolding like SLGO. He describes a three-layer defense against prompt injection attacks combining alignment research, a mechanistic-interpretability-based classifier that monitors neurons identified by Chris Olah’s team, and an auto-mode classifier — a combination that has effectively eliminated demonstrable prompt injection in testing.
Cherny reveals that Claude Code’s system prompt was reduced by more than 80% with Opus 5, because the model’s reasoning has advanced past the point where explicit behavioral corrections are needed. He also shares a concrete capability milestone: the Bun JavaScript runtime team used Claude to rewrite portions of the Bun codebase from Zig to Rust in a single prompt, a task that would have been impractical with earlier model generations. The key enabler was Bun’s extensive test suite, which gave the model a reliable exit criterion.
For founders and builders, Cherny argues that most people are under-utilizing current models by providing overly prescriptive step-by-step instructions rather than high-level goals with defined guardrails and exit criteria. He frames this as a significant product overhang — model capabilities have outrun the way most users are actually deploying them — and encourages experimentation with Claude Code’s undocumented “simple mode” (CLAUDE_CODE_SIMPLE=1) to explore baseline model behavior without system prompt scaffolding.
📺 Source: Y Combinator · Published July 27, 2026
🏷️ Format: Interview
