Opposable's AI phone agent takes 50 seconds to toggle dark mode

An AI with hands on your phone sounds like a shortcut and reads less like one once you time it. Reval Labs released Opposable, an Android app that reads the phone screen through accessibility services and taps and types in other apps, under the control of an AI model. The company's own blog post this week carries the measurements that will decide whether anyone keeps using it.

Writing on getopposable.com, the Reval Labs team describes two paths. In one, a desktop AI client that supports MCP, such as Claude Code, ChatGPT or Codex, drives a connected Android device. In the other, an on-device agent runs with no desktop at all, with a 2B vision-language model called GUI-Owl 1.5 2B inside the Android app itself.

What the vendor's own numbers show

Three worked examples sit on Opposable's site. Writing in The Register, Brandon Vigliarolo timed them. Changing the system to dark mode and setting a 5-minute timer takes 50 seconds and 17 tool calls over Claude Code with MCP. Looking up when Apollo 11 launched and who was on board takes close to two minutes with a run cost of $0.44. Opening the clock app and switching to the stopwatch tab, under the on-device agent and with no desktop in the loop, takes 96 seconds.

The post is more direct on accuracy. On eight held-out tasks that the model had never been tuned on, run twice each on an emulator, giving the whole task to the model succeeded 4 times out of 8 in the first round and 3 times out of 8 in the second. Giving the model one subgoal at a time and checking each claimed success against the accessibility tree before the next step raised that to 6 of 8 in both rounds. The post is clear that the second number came with the agent being guided step by step.

Where the speed does come from

The faster figures in the post are at the step level, not the task level. On a MacBook with an M2 CPU running four threads, a full step takes 7.3 seconds today, down from 10.6 seconds, with the vision encoder taking 3.6 seconds, the prefill 2.5 seconds and the decode 1.1 seconds. The accuracy of that step is 27 of 36 screens correct.

A Galaxy S25 moves the vision encoder onto its neural processing unit at FP16 precision, which cuts the encoder step from 3,900 milliseconds on an M2 CPU to 236 milliseconds. First load takes 72 seconds and later loads 0.39 seconds. The GUI-Owl 1.5 2B model is a Qwen3-VL fine-tune quantised to 4 bits, with the language half at 1.05 gigabytes and the vision encoder at 445 megabytes.

Vigliarolo also points out that 40 common tasks, setting an alarm, sending a text, switching Wi-Fi, run as fixed code with no AI model involved. He quotes a Reval Labs warning on the Opposable site: "whoever holds the token can operate the phone, so treat it like a password."

Source

• Computer use on the phone itself: a 2B model that operates Android apps by Reval Labs
• Reach new heights of laziness by burning tokens to let an AI tap your smartphone screen for you by Brandon Vigliarolo, The Register