Open-source Buddy benchmark tests which AI model uses a real Android phone fastest
3amae404 · reddit · 2026-10-18
A developer built Buddy, an open-source Android app where Claude Code, Codex, or OpenCode operates a real phone, and benchmarked 7 models on a OnePlus 12 across 5 everyday tasks (2 runs each, apps reset, no approvals). Sonnet 5.5 was fastest at 25s (10/10 correct); Opus 5.5 took 27s with fewest steps; GPT-6 Luna hit 29s but only 7/10 correct; free OpenCode models Space Bunny and Big Pickle finished in 43s and 112s. Surprises: Opus skipped Maps UI entirely via a directions deep link (15s), some models went straight to wttr.in for weather, and Luna twice reported the daily high instead of current temperature. Data and method are open-sourced.
Related event: Buddy Tests AI Agents Controlling Real Android Phones(2 posts)→
More from coding & agent
- Websites are locking out AI agents: Yelp charges, Delta and JPMorgan block outright — gaganghotra_ · 2026-10-18
- 7-item checklist to stop your agents and automations from failing silently — shannonpieter · 2026-10-18
- Weekly puppet sketch show written, directed and animated entirely by Claude agents — Slackluster · 2026-10-18
- AgentCave puts your scattered Macs and GPU boxes in one virtual room — noahsolomon · 2026-10-18
- Letting LLMs maintain the OpenAPI spec: perfection isn't a prerequisite — charliermarsh · 2026-10-18
- User Pleads With Anthropic: One Continuous AI Chat With Full Memory — khademinori · 2026-10-18