Build an Android app
Pick the best app-building model from long-horizon Android work, engineering judgment, general long-horizon engineering, and mergeability.
Right now Claude Sonnet 5.5 is first on 2 of 3 boards it enters.
Sorted by compositeSynced 2026-10-10
| Claude Sonnet 5.5 | 95 | 3/4 | 1 | 1 | — | 4 |
| Claude Opus 5.5 | 92 | 3/4 | 2 | 2 | — | 2 |
| GPT-6 Astra | 79 | 4/4 | 4 | 5 | 1 | 3 |
| GPT-6.1 Sol | 75 | 3/4 | 3 | 3 | — | 5 |
| Claude Fable 5.1 | 64 | 3/4 | 5 | 4 | — | 7 |
| Gemini 4 Argon | 52 | 3/4 | 6 | 8 | — | 12 |
| Qwen3.8 Max | 46 | 3/4 | 7 | 6 | 16 | — |
| Gemini 3.7 Flash | 45 | 2/4 | 11 | — | 12 | — |
| Gemini 3.8 Flash | 44 | 3/4 | 10 | 15 | 2 | — |
| Kimi K3 | 44 | 4/4 | 8 | 13 | 8 | 19 |
| GPT 6 Luna | 40 | 2/4 | 9 | — | — | 16 |
| ?Beam 1.0 | — | 1/4 | — | — | — | 25 |
| Claude Fable 5 | — | 1/4 | — | — | 5 | — |
| Claude Haiku 5.5 | — | 1/4 | — | — | — | 9 |
| Claude Opus 4.8 | — | 2/4 | — | — | 15 | 21 |
| Claude Opus 5 | — | 2/4 | — | — | 3 | 11 |
| Claude Sonnet 4.6 | — | 1/4 | — | — | 27 | — |
| Claude Sonnet 5 | — | 1/4 | — | — | 18 | — |
| DeepSeek V4 Flash | — | 1/4 | — | — | 20 | — |
| DeepSeek V4 Pro | — | 1/4 | — | — | 14 | — |
Suggested build
Why it matters
The intervals overlap
First is 32.7% (18.7–47.3). Second is 28.0% (13.3–42.0). The gap is small, so rank alone is a weak reason to switch.
Fourth is under half the cost
The leader's run costs $215.4. GPT 6 Sol is fourth at 18.0% and $93.4, under half that cost.
1.0 is no longer updated
The 1.0 standard-task table is still the full 32 rows, marked archived on the benchmark page. The scenario and the home card use only 2.0 long-horizon.