DeepSeek V4.1 Flash 1017 Bug-Hunt Scores: Harness Matters More at Higher Effort
PawelHuryn · x · 2026-10-19
Pawel Huryn published full Bug Hunt Bench results for DeepSeek V4.1 Flash 1017 across 3 agent harnesses (Claude Code, DeepSeek Harness, Pi) and 3 effort levels. Best run: Max effort + DeepSeek Harness at 42.7/105 mean ($1.84/run); same effort on Claude Code averages just 39.1. Low-effort runs cluster at 34 regardless of harness, with costs as low as $0.46. Takeaway: the harness plays a bigger role at higher reasoning effort.
- 17 runs total, each 33-62 minutes, blind-judged diffs
- Full per-run score table included
Related event: Bug Hunt Bench: Same Model, Harness Swap Shifts Score by 3.6 Points(6 posts)→
More from coding & agent
- Dev community notices Codex has 'suddenly got really good' at coding — enggirlfriend · 2026-10-19
- Claude Code lead gets hounded for resets on the street as users say quota dies in 3-10 hours — umesh_ai · 2026-10-19
- Agent auth is still broken: passkeys bottleneck remote AI workflows — curious_vii · 2026-10-19
- Scale CEO pushes Muse for coding as early users call it surprisingly pleasant — alexandr_wang · 2026-10-19
- AI agents cost 20x more than conventional software, warns architect's enterprise breakdown — DavidLinthicum · 2026-10-19
- Indie dev shares UI iteration workflow: three mock options for every visible change — SeanOliver · 2026-10-19