Code with Claude 2026: Half of Devs Ship PRs They Never Read
At Anthropic's May event, nearly 50% of attendees reported shipping Claude-written pull requests without reading the code first. SWE-bench is at 87%. The model is no longer the bottleneck — discipline is.
NeuroX AI · May 23, 2026
At Code with Claude 2026 last week, Anthropic revealed a number that should land like a cold shower: nearly 50% of attendees shipped pull requests written entirely by Claude — without reading the code first. Same week, Claude Opus 4.7 hit 87% on SWE-bench, up from 62% on Sonnet 3.7. Boris Cherny, Claude Code's leader, put it plainly: "The default is now 'I'm going to have Claude prompt itself.'"
This is a different kind of inflection than capability benchmarks usually signal. The model cleared the bar for autonomous shipping. The open question is whether the teams using it have cleared theirs.
Shipping unread code fast is a productivity multiplier — until the third week, when the accumulation of unreviewed abstractions, silent regressions, and absent test coverage forces a rewrite. The orgs in Anthropic's own case studies — Rakuten with 99.9% accuracy on a 12.5M-line rewrite, Zapier with 800+ governed agents — didn't skip the discipline layer. They built it first and let the model run inside it.
Anthropic also shipped "Dreaming" — a Managed Agents feature that consolidates an agent's memory between sessions, pruning duplicates and stale context automatically. It's a sign the platform is building toward agents that improve themselves. That only works if the agent's operating environment is well-defined to begin with.
The teams who'll benefit most from the next capability jump are the ones with evals, scoped tools, and review gates already in place. See how we build that layer in 30 days →