Google's Android Bench ranks Gemini 4 Argon sixth for Android coding

Google's own ranking of AI models for Android coding has an awkward result for Google. In the latest Android Bench results, as of 26 October, Gemini 4 Argon sits behind several Claude and GPT-6 models ^[1].

What the benchmark shows

Android Bench 2.0 measures how well AI models and coding agents handle 30 tasks, and its pass rate is the average percentage of those tasks successfully resolved across all runs ^[1]. The results now also show the reasoning level used for each model ^[1].

Claude Sonnet 5.5 tops the table with a 40.0% pass rate and an 80.9% completion rate, followed by Claude Opus 5.5 at 37.3% ^[1]. OpenAI's GPT 6.1 Sol is third at 29.3%, then GPT 6 Astra at 28.0% and Claude Fable 5.1 at 22.7% ^[1]. Gemini 4 Argon, run at "high" reasoning, scores 21.3% with an 81.8% completion rate, in sixth place ^[1]. 9to5Google reported the model as fifth, but Google's own table lists it sixth ^[1][2].

Cost and speed

The table also lists the average cost of a full benchmark run, and the spread is large. Sonnet 5.5 costs $1,574.90, while GPT 6.1 Sol costs $99.50 ^[1]. Argon's run cost $267.60 and took an average of 52.8 hours, against 26.6 hours for GPT 6.1 Sol ^[1]. 9to5Google called GPT 6.1 Sol "the most interesting addition" because of that price gap ^[2].

Reading it carefully

A few cautions apply. The results are for a single benchmark of 30 tasks, and Google's table gives a statistical range for each score. Argon's range runs from 10.0% to 33.3%, which overlaps heavily with Fable 5.1 (10.7% to 36.0%) and even touches the lower end of GPT 6.1 Sol (15.3% to 42.7%) ^[1]. Sonnet 5.5's range is 24.0% to 56.0% ^[1].

The models are also paired with different agent software. Argon was run with Google's antigravity-sdk, the Claude models with claude-code and the OpenAI models with codex ^[1]. So the table compares model and agent combinations, not models alone.

Completion rates, which give credit for partial progress, are far closer: Argon's 81.8% sits near Fable 5.1 (82.4%) and GPT 6 Astra (82.2%) ^[1].

Background

9to5Google notes that Android Bench was introduced earlier this year to help Android developers see which models suit Android coding, and that cost was added so developers can weigh price against quality ^[2]. It also recalls an earlier report in which internal voices claimed Argon "struggles to handle certain coding tasks", though Google stood by the model's capabilities ^[2].

What it means for businesses

If your company builds Android apps, this is useful evidence but not a verdict. The gap between top and bottom on cost is much bigger than the gap on completion rate, so the cheapest capable option may suit routine work, while the top scorer may justify its price on harder migrations. Test two or three models on your own code before committing, because a public benchmark cannot reflect your codebase. For Google, the result is a reminder that it publishes a ranking in which its own newest model does not lead.

Source: Google — Android Bench