Android Bench July update moves to the Harbor framework and re-baselines scores

What happened on this date

On 2026-07-08, the Android Developers Blog published a post by Zoe Lopez-Latorre on the July update to Android Bench. Android Bench is the LLM leaderboard Google introduced in March for real-world Android development tasks, such as Jetpack Compose migrations, wearable networking and platform API updates. The changes: evaluation moved to the Harbor framework with an updated benchmarking agent (previously mini-swe-agent v1 adapted for Android), making results easier to reproduce; all models were re-run, with historical scores kept in an archive; and eight models were added. The post reports a top score of 84.5, and developers can now contribute tasks and evaluation results through GitHub repositories.

What it means for assessing AI-assisted cloning

Android Bench measures how well models complete Android development tasks; it does not measure reverse engineering or copying, so its scores should not be read as a cloning capability. Indirectly, though, it shows that implementing common UI, migration and platform-integration work with models is becoming more feasible. For independent developers, this means cloning risk comes not only from taking code but from re-implementing what is visible. If an app's value lies mainly in its interface and client-observable flows, no amount of binary protection prevents someone rebuilding it. Code protection raises the cost of understanding and lifting a specific implementation; it does not stop others shipping a functionally similar product.

Checks to run for your own product

First, list your product's features and mark which could be rebuilt from the interface and public behavior alone, and which depend on server-side data, models, rules or user network effects. Second, keep the genuinely differentiating logic on the server, with authentication, rate limiting and anomaly monitoring on its APIs, so visible features are separated from capabilities that cannot be copied. Third, keep only necessary local logic on the client and re-check entitlement and billing decisions server-side. Fourth, periodically search app stores for apps with similar interfaces or names and keep evidence for complaints. Fifth, if your team uses models for development, public benchmarks like this can inform selection, but verify against your own tasks rather than relying on a leaderboard.

Your checklist

• Separate features rebuildable from the UI from those relying on server capabilities
• Keep differentiating logic server-side with authentication and rate limits
• Re-check entitlement and billing on the server
• Periodically search stores for look-alike apps and keep evidence