Google's own AI places fifth in its Android coding test

Google's own AI model did not do well on Google's own Android coding test. In the October 26 update to Android Bench, Gemini 4 Argon placed fifth, well behind the latest models from Anthropic (Claude) and OpenAI (GPT).

What the list shows

Android Bench launched early this year. It lets app developers see which AI models help most with Android coding, using a test built by Google. Claude Sonnet 5.5 leads with a 40% pass rate, meaning the share of test tasks it passed. Claude Opus 5.5 is second.

The update also added the reasoning level used for each model, plus results for GPT-6 Sol, GPT-6.1 Sol, Claude Sonnet 5 and Gemini 4 Argon.

Price matters too

The list now includes average cost. GPT-6.1 Sol passed 29.3% of tests at a cost of $99.50. The top-ranked Claude model cost $1,574.90. A lower score at a far lower price may suit some projects.

A report shortly after Gemini 4 Argon was announced quoted internal voices saying it "struggles to handle certain coding tasks." Google stood by the model's capabilities.

Why it matters

If you pay developers to build an app, ask which AI tools they use. Quality and price vary a lot. The material does not say how much any of this changes real app costs or deadlines, and it does not include a new response from Google on the ranking.