Google's Own Benchmark Puts Gemini 4 Argon Fifth For Android Coding

Claude Sonnet 5.5 now sits at the top, with a 40% pass rate and an 80.9% completion rate. Claude Opus 5.5, added to the list in early October, holds second place.

Gemini 4 Argon comes in fifth, behind Claude Fable 5.1, which posted a 21.3% pass rate, an 81.8% completion rate and an average cost of $267.70 per task. Google's post didn't break out Argon's own pass-rate figure, only that it finished behind that result on the leaderboard.

Price tells its own story here. GPT-6.1 Sol posted a 29.3% pass rate and an 82.8% completion rate, both short of Claude Sonnet 5.5's numbers, but it did it for an average cost of $99.50 per task. Claude Sonnet 5.5's chart-topping run cost $1,574.90 on average. For a developer choosing between the two, that's a roughly sixteen-fold price gap for a nine-point pass-rate difference, and it's the kind of tradeoff Android Bench was built to surface.

The bigger embarrassment is that Gemini 4 Argon is Google's model, running on Google's own benchmark, and it isn't close to the top of either the quality list or the value list. 9to5Google reported that unnamed internal voices at Google said the model "struggles to handle certain coding tasks." Google has stood by Argon's capabilities, according to the same report, but hasn't offered a rebuttal of the specific benchmark numbers it published itself.

Android Bench doesn't claim to measure general intelligence or everyday chatbot usefulness. It is narrow by design, built around the kind of code an Android app developer actually writes, which is part of why its results carry weight with that audience even when they're unflattering to the company running the test.

Whether Gemini 4 Argon improves in a future update, or whether Google ships a revised model to answer the complaints reportedly circulating inside the company, isn't something Google has said yet. The leaderboard, for now, stands as published.