HackerRank-Bench
Code
HackerRank's private benchmark of long-horizon engineering work on complete software systems: greenfield builds, brownfield changes, migrations, identity and provisioning, frontend and privacy-sensitive data work, all written from scratch. Every model runs in the OpenHands agent harness from the same saved workspace, and hand-written hidden verifiers award each run the weighted share of behavioral checks it passes. The score is the mean over tasks and three runs per task; the task count is not disclosed.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 51.2 |
| Gemini 4 Argon | 49.7 |
| Claude Fable 5.1 | 47.5 |
| GPT-6 Astra | 43.5 |
| Claude Opus 5 | 42.6 |
| GPT-5.6 Sol | 38.4 |
| Kimi K3 | 38.3 |
| Grok 4.5 | 34.2 |
| Gemini 3.7 Flash | 28.5 |
| Sonnet 5 | 28.0 |
| Grok 4.6 | 27.9 |
| GPT-5.6 Terra | 26.1 |
| GPT-5.6 Luna | 26.1 |
| GLM-5.2 | 23.7 |
| MiniMax M3 | 21.1 |