Google's Android Bench leaderboard, updated Oct. 26, ranks Gemini 4 Argon sixth of 11 on its long-horizon tasks. Argon passes 21.3% of them. Claude Sonnet 5.5 leads at 40.0%, with Claude Opus 5.5 at 37.3% and GPT 6.1 Sol at 29.3%. The changelog lists Argon's results as new that day. 9to5Google counts it fifth; the live table shows Claude Fable 5.1 at 22.7% just above it.
The numbers
Long-horizon leaderboard, with the reasoning level and agent each model ran under:
-
- Claude Sonnet 5.5 (max, claude-code): 40.0% pass, 80.9% completion, $1574.9 average cost
-
- Claude Opus 5.5 (max, claude-code): 37.3% pass, 85.6% completion, $1150.3
-
- GPT 6.1 Sol (max, codex): 29.3% pass, 82.8% completion, $99.5
-
- GPT 6 Astra (high, codex): 28.0% pass, 82.2% completion, $375.7
-
- Claude Fable 5.1 (high, claude-code): 22.7% pass, 82.4% completion, $492.6
-
- Gemini 4 Argon (high, antigravity-sdk): 21.3% pass, 81.8% completion, $267.6
- Next: Qwen3.8 Max at 12.5%, Kimi K3 at 12.0%, GPT 6 Luna at 10.7%, Gemini 3.8 Flash at 8.0%, Gemini 3.7 Flash at 7.3%
Argon's confidence range is 10.0% to 33.3%. That overlaps Fable 5.1 and GPT 6 Astra, so those three places are close. Sonnet 5.5's range starts at 24.0%, above Argon's central score.
What the test covers
Google says the long-horizon set is 30 tasks, run five times each: 10 app creations, 13 migrations, 5 new features and 2 app conversions. Examples on the page include building a food-delivery app, converting the Habo Flutter app and the BlueWallet React Native app to Android, and moving Bitwarden and Signal code between libraries.
Pass rate is the share of tasks resolved across runs. Completion rate measures how close a run got even when it failed. Cost is the average for a full benchmark run, and latency is the average time to solve the 30 tasks. Argon's 52.8 hours is slower than GPT 6.1 Sol's 26.6 and Fable 5.1's 22.2, and faster than Sonnet 5.5's 70.7.
The Oct. 26 update also added a column for the reasoning level each model used, and archived the previous Claude Sonnet 5 and GPT 6 Sol results.
Why it matters
Argon's completion rate, 81.8%, sits within about a point of Fable 5.1 and GPT 6 Astra, yet its pass rate is lower. On these tasks it gets close to a working app about as often as they do but finishes the job less often, and a half-working migration still has to be fixed by hand. Cost does not close the gap either: GPT 6.1 Sol passes more tasks for $99.5 against Argon's $267.6.
Google runs this benchmark and Argon is its own model, so the ranking is not a rival's claim. Each model also ran under its own agent tool, so the table compares model-and-tool pairs, not models alone. The older, easier 1.0 set of 100 tasks lists no Argon result yet.
What we checked
We read the leaderboard, task counts, metric definitions and changelog on Google's page on Oct. 27. Every score above comes from it.
Related: Gemini 4 Argon reaching the Gemini app and GPT-6.1 Terra pricing and benchmarks.
