Opposable launches an on-device phone agent, 7.3 seconds per tap

A phone agent that writes "give your AI thumbs" on its home page needs to say how often the thumbs land where they are told. Opposable, a product from Reval Labs, released its Android app on 21 October 2026 together with a technical post that gives the numbers behind every step. The app reads the Android screen through the accessibility service, taps and types with a 2-billion-parameter vision-language model, and runs that whole loop on the phone when no desktop agent is attached. Opposable also connects to Claude, ChatGPT or Codex over MCP from a paired computer.

The timings

On a Mac M2 at 4 threads, Opposable says one step takes 7.3 seconds, down from 10.6 before its own CPU work. A step is three stages: the vision encoder turns the screenshot into tokens in 3.6 seconds, the model reads the prompt in 2.5, and writes the action in 1.1. On a Galaxy S25 with a Snapdragon 8 Elite neural processor, the encoder alone drops from 3.9 seconds on the laptop to 236 milliseconds on the phone's NPU, 16 times faster by Opposable's own measurement. The model is GUI-Owl 1.5 2B from Alibaba's Mobile-Agent team, run with llama.cpp at 4-bit, 1.05 GB for the language model and 445 MB for the vision encoder.

The success rate

Opposable also published two accuracy rounds on eight tasks the model had not been tuned on. Giving the model the whole task at once succeeded 4 of 8 and 3 of 8 across two rounds. Giving it one subgoal at a time and checking each claim against the accessibility tree before moving on succeeded 6 of 8 in both rounds. Those are Opposable's own numbers on an Android emulator. The company's post also lists limits: only Snapdragon has an NPU build, Pixel and MediaTek phones fall back to the GPU or CPU, and the on-device model is weaker than a cloud model on apps it has not seen.

What a team can take from this

Two readings. One, an on-device agent costs no tokens and keeps the screen on the phone, which is the pitch, and 236 milliseconds for a vision encoder on a 2026 Android is a real finding other teams building for mobile can plan against. Two, 3 to 4 of 8 tasks succeeding on a first try is below the success rate a daily task needs, and steps that need the 640-token input size still encode on the CPU because that size did not compile for Opposable's NPU target. A product that gives an agent control of a user's phone has to decide what happens when four of its eight tries fail, before that turns into a support ticket.

Source

Primary: Opposable blog, "Computer use on the phone itself". Press: The Register, Brandon Vigliarolo.