Claude Code CLI Outperforms Every GUI Agent — Researchers Say the Phone Screen Might Be Obsolete
Claude Code Opus 4.7 achieves 71.8% on AndroidWorld, surpassing all GUI baselines.
A new study argues that mobile agents don't need to interact with phone screens via GUI—they can use the device’s command-line interface (CLI) instead. Researchers evaluated three coding agents (Claude Code, Terminus-2, mini-swe-agent) on AndroidWorld and MobileWorld without any mobile-specific training, comparing them against three reproducible GUI baselines (GUI-Owl-1.5-32B, MAI-UI, Qwen3-VL-32B). Claude Code Opus 4.7 reached 71.8% on AndroidWorld and 51.9% on MobileWorld, outperforming every GUI baseline (best: 69.3% and 43.2%). Oracle CLI solutions hit 88.8% and 86.3%, showing substantial headroom.
The CLI advantage is not just performance—efficiency is key. CLI agents completed tasks in significantly fewer steps (10.7 vs 18.6). To cover real-world intents beyond GUI scope, the team created the CLI-Advantage Task Suite (45 templates across five categories: bulk operations, multi-condition filtering, aggregation, cross-app workflows, hidden device state). In every category, CLI agents beat all GUI baselines. The authors will open-source agent implementations, oracle solutions, and evaluation infrastructure, inviting further research into CLI-first mobile agent design.
- Claude Code Opus 4.7 scored 71.8% on AndroidWorld, beating the best GUI baseline (69.3%) and 51.9% on MobileWorld (vs GUI best 43.2%).
- CLI agents used 10.7 steps per task on average, nearly half the 18.6 steps required by GUI agents.
- New CLI-Advantage Task Suite (45 templates) proves CLI dominance in five categories critical for everyday mobile automation.
Why It Matters
Mobile AI agents can now bypass screen simulation, using CLI for faster, more capable task automation on phones.