Build an Android app
Pick the best app-building model from long-horizon Android work, engineering judgment, and general long-horizon engineering.
Right now Claude Opus 5.5 is first on 2 of 2 boards it enters (1 tied).
Sorted by compositeSynced 2026-10-10
| Claude Opus 5.5 | 97 | 2/3 | 1 | 1 | — |
| GPT-6 Astra | 86 | 3/3 | 2 | 3 | 1 |
| Claude Fable 5.1 | 69 | 2/3 | 3 | 5 | — |
| GPT 6 Sol | 56 | 1/3 | 4 | — | — |
| Kimi K3 | 50 | 3/3 | 6 | 8 | 8 |
| Qwen3.8 Max | 50 | 3/3 | 5 | 6 | 16 |
| Gemini 3.8 Flash | 45 | 3/3 | 8 | 12 | 2 |
| GPT 6 Luna | 42 | 1/3 | 7 | — | — |
| Claude Sonnet 5 | 39 | 2/3 | 10 | — | 18 |
| Gemini 3.7 Flash | 36 | 1/3 | 9 | — | — |
| Claude Fable 5 | — | 1/3 | — | — | 5 |
| Claude Opus 4.8 | — | 1/3 | — | — | 15 |
| Claude Opus 5 | — | 1/3 | — | — | 3 |
| Claude Sonnet 4.6 | — | 1/3 | — | — | 27 |
| Claude Sonnet 5.5 | — | 1/3 | — | 2 | — |
| DeepSeek V4 Flash | — | 1/3 | — | — | 20 |
| DeepSeek V4 Pro | — | 1/3 | — | — | 14 |
| Gemini 3.1 Pro Preview | — | 1/3 | — | — | 28 |
| Gemini 3.5 Flash | — | 1/3 | — | — | 25 |
| Gemini 3.6 Flash | — | 1/3 | — | — | 23 |
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1.0 is no longer updated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.