Breken's AI bug finder runs your app, beats Claude Code
Why it matters
Startup Breken launched on Oct 21 with what it calls production engineering: instead of reading the diff like AI code review, it rebuilds your entire software on its own machines and grows a production history at 4,000x speed, with outside services answered inside an isolated environment. Synthetic users and agents work the product (impatient admins, users with two tabs open, agents acting on inboxes) while Breken changes one condition at a time, such as a retry, a lost response or a clock jump. A failure counts only if it reproduces every time, and you get the exact trace plus a draft pull request. On its ESCAPES benchmark of 327 real, user-reported bugs from 33 open-source products, 235 were already present in the code as of July 1, 2026; Breken caught 76 of them with 20 false alarms, versus 37 for Claude Fable 5.1 in Claude Code (59 false alarms), 33 for Claude Opus 5.5, 28 for GPT-6 Astra in Codex CLI and 9 for Semgrep. Breken says 71% of these bugs only appear when the product runs, with real data, users, failures or time. Caveat: the benchmark is vendor-run; Breken built it and ran every system itself, though the bugs, prompts, judge prompts and labels are public.
Read the source: Breken (opens in a new tab)
All hits from 23 Oct 2026More CodingGet the daily hits by email