Google just changed how it measures which AI models are best at writing Android app code, and the update has shuffled the rankings developers use to pick their tools. The company’s Android Bench leaderboard, which launched in March, now runs on a new testing system called Harbor. Google says this replaces the older, more generic testing tool it used before, and gives a better read on how models perform on real Android tasks, like updating old code to Jetpack Compose or handling wearable device networking.
New models shake up the top of the list
Since the testing tool changed, Google ran every model through it again. Eight new models were added to the leaderboard, including Claude Fable 5, Claude Sonnet 5, Claude Opus 4.8, GLM 5.2, Kimi K2.7 Code, MiniMax M3, Qwen 3.7 Plus, and Qwen 3.7 Max.
Android Bench
Claude Fable 5 now sits at the top with a score of 84.5 percent, followed by GPT 5.5 at 80.2 and Claude Sonnet 5 at 76.2. Among free, open-weight models, GLM 5.2 leads with 72.2 percent, ahead of Kimi K2.7 Code at 70.4. Google’s own Gemini 3.1 Pro is placed fifth with a score of 73.7 percent. If you refer to Android Bench to pick the right model for your coding work, you should head over to the Android Bench website and check the refreshed leaderboard.
Developers can now submit their own test tasks
When Google launched Android Bench in March, it published its testing methodology and test harness on GitHub for transparency. Now, it’s taking that open approach further by letting developers contribute their own Android development tasks to the benchmark. Developers can also run the tests themselves with their preferred models and share the results with the community.
The rankings are most likely to change as more new models show up, like OpenAI’s recently released GPT-5.6 Sol, Terra, and Luna. If you’re picking an AI tool for Android coding, treat the current leaderboard as a snapshot, not a permanent ranking, and check back before you commit to one for your next project.
