Google identified top AI models for creating Android apps—Gemini fell behind GPT

Google updated the Android Bench ranking

Google has once again revised its list of top AI models for Android development. The new lineup includes many open weights, as well as detailed statistics on token usage and cost per model.

What’s new in the tests
- Average latency – time spent solving 100 tasks over ten runs.
- Total tokens – average token consumption per run (sum across ten runs).
- Cost – U.S. dollar expenses for running a single benchmark.

These metrics make results more transparent and allow comparisons not only of speed but also of economic efficiency of the models.

Top models
Position | Model | Key data
1 | OpenAI GPT 5.5 (May 18) | About 2 % faster than Gemini 3.1 Pro, but more expensive to run
2 | Google Gemini 3.1 Pro | Most economical among leaders; almost twice cheaper than GPT 5.5
3 | OpenAI GPT 5.4 | On par with Gemini 3.1 Pro, but slower than GPT 5.5
4 | GLM 5.1 (open model) | Best result among open weights

What to expect next
Google recently unveiled a new version of Gemini – Gemini 3.5 Flash, and soon a more powerful Gemini 3.5 Pro will appear. It will be interesting to see how they compete with the current leader GPT 5.5 and how cost and speed metrics change.

Thus, the Android Bench ranking now not only shows which model is faster but also allows assessment of the economic aspect of working with AI in Android app development.