Google’s October 26, 2026 Android Bench 2.0 leaderboard placed Gemini 4 Argon sixth, with a 21.3% pass rate. Claude Fable 5.1 ranked fifth at 22.7%, though the two models’ displayed performance intervals overlap. The results cover Android engineering tasks, not coding ability across every programming domain.
Gemini 4 Argon’s place in Google’s results
Google’s Android Bench leaderboard lists Gemini 4 Argon (high) with the antigravity-sdk agent at 21.3% pass rate, 81.8% completion rate, 52.8 hours average latency, and an average cost of $267.6 per full benchmark run. The results were labeled latest as of October 26.
Claude Fable 5.1, immediately above Argon, recorded a 22.7% pass rate. Their displayed intervals overlap: 10.7% to 36.0% for Fable and 10.0% to 33.3% for Argon. The table sets out an order, but those figures do not establish that the small pass-rate gap is statistically significant.
| Model and agent | Pass rate | Displayed interval | Completion rate | Average latency | Average cost per full run |
| Claude Sonnet 5.5 (max), claude-code | 40.0% | 24.0% to 56.0% | 80.9% | 70.7 hours | $1,574.9 |
| Claude Opus 5.5 (max), claude-code | 37.3% | 22.0% to 52.0% | 85.6% | 56.6 hours | $1,150.3 |
| GPT 6.1 Sol (max), codex | 29.3% | 15.3% to 42.7% | 82.8% | 26.6 hours | $99.5 |
| GPT 6 Astra (high), codex | 28.0% | 13.3% to 42.0% | 82.2% | 7.9 hours | $375.7 |
| Claude Fable 5.1 (high), claude-code | 22.7% | 10.7% to 36.0% | 82.4% | 22.2 hours | $492.6 |
| Gemini 4 Argon (high), antigravity-sdk | 21.3% | 10.0% to 33.3% | 81.8% | 52.8 hours | $267.6 |
What Android Bench 2.0 tests
Google’s methodology describes 30 long-horizon Android engineering tasks: 9 app-creation tasks, 13 migrations, 6 new-feature tasks, and 2 cross-platform app conversions. Each task is run independently five times to account for variation in model behavior.
A pass requires a complete solution with a perfect score, all functional tests passing, full visual compliance, and no constraint violations. The completion rate is a continuous score that captures progress on runs that fall short of a pass, using functional, regression, requirements, and visual checks with penalties for violations. That helps distinguish a nearly finished task from one that made little progress, even though both may count as non-passes.
How to read cost and latency
The listed cost is an average per full benchmark run, not a subscription price or a cost per successful task. For example, GPT 6.1 Sol’s listed $99.5 run cost is $1,475.4 below Claude Sonnet 5.5’s $1,574.9, while its pass rate is 10.7 percentage points lower: 29.3% versus 40.0%.
That cost gap does not show which model spends less per solved task. Google cautions that early failures can lower total resource use, so a lower gross cost or shorter runtime can reflect less work completed. Latency also includes network transit, and cost reflects provider API prices at the time of the benchmark run.
What the ranking can tell you
Android Bench 2.0 evaluates a defined set of Android development tasks using model-and-agent configurations. Its rankings describe performance on that test set; they do not establish how the same systems compare on iOS, web development, or programming tasks in general.
