Anthropic's 400K Claude Code sessions report: managers beat engineers on verified success

2026-06-22 群聊日报

Anthropic analyzed ~400K real Claude Code sessions and found managers scored highest on verified success—tasks requiring git commits, merged PRs, and passing tests—beating software engineers. The group flagged selection bias: managers chase aha moments, not corner cases. Sakana AI launched Fugu, a multi-model orchestration product hitting SWE Bench Pro 73.7, explicitly marketed as free from US export controls. OpenCode's public dashboard shows 136K DAU with DeepSeek at 53.6% share—domestic Chinese models dominate. On the practical side, the root cause of Gemini 3.5 Flash output truncation was traced to thinking monologues consuming tokens before the response could start; increasing the output limit helps but remains unstable.

Why it matters: Anthropic's research on 400K real Claude Code sessions shows managers scoring higher than engineers on verified success, with selection bias already flagged in the chat. HKR all hit, but this is a second-hand digest from a daily roundup rather than a primary source, so 72 at t...