The constraints were strict to keep the benchmark honest: one single prompt, no follow-up corrections, and a completely blank Expo project. If the code didn't run on a physical device, it was a fail. I tested GPT-5.6 Sol High, Claude Fable 5, and Kimi K3 High.
Here is the exact prompt used for the deep dive:
Recreate the Robinhood app screens shown in the attached video as a React Native
(Expo) app. Build three screens: Discover, Token detail (scrubbable chart with
price + Buy reveal), and the Swap/amount keypad screen.
Priority is fidelity of motion and feel, not features — match the real app's
easing, timing, gesture response, haptics, and typography (numbers should
animate digit-by-digit). Use Reanimated + Gesture Handler, and Skia for the
chart. Mock all data locally; no backend.
Watch the video closely and match what you see frame for frame. Tell me which
interactions you couldn't reproduce and why.
The Performance Breakdown #
GPT-5.6 Sol High took the win here. It was the only model that truly understood a "motion brief" as a requirement for feel rather than just layout.
Tokens used: 2.0MEstimated cost:~$15** Time to first render:32 minutes Dependencies added:**3
The result was surprisingly accurate. It nailed the odometer-style rolling digits on the price ticker—a detail most LLMs fail at by defaulting to a simple cross-fade. The Skia chart was actually scrubbable, and the layout was nearly identical to the reference.
However, there is a technical caveat. Like the other models, GPT-5.6 handled the ticker and chart redraws via React state instead of utilizing UI-thread worklets. This introduces a "JS-thread tax," meaning the interactions aren't as buttery smooth as the actual Robinhood app. If you're looking for a professional deployment, you'll still need to manually optimize those animations to move them off the main thread.
Despite that, it outperformed the others in speed and cost, hitting the mark in half the time of the runner-up. For anyone building a complex AI workflow for mobile UI, this proves that the latest reasoning models are getting significantly better at interpreting visual timing and easing from video references.
Next Why is everyone still trying to force Regional Prompter onto →
a guide to making money with AI, with plenty of directly applicable cases.