SaaSBench (Engineering): leaderboard
SaaSBench (Ren et al., 2026) tests long-horizon software engineering on 30 open-source SaaS categories in six domains. Coding agents build and deploy an enterprise SaaS application from scratch from a product requirements document of about 4,400 lines in a Docker workspace, then a dependency-aware graph of HTTP, database, container, browser and rubric checks scores it. Mean task score in percent over the 30 tasks, one rollout each, with each backbone run as a separate OpenHands or Claude Code row; higher is better.
Metric: Task score (%; pass@1, one rollout per task: mean over the 30 tasks of the share of each task's validation-DAG max score earned, with deterministic HTTP, database, container and browser checks plus Claude Sonnet 4.5 rubric nodes; the agent builds and deploys an enterprise SaaS system from a PRD in a Docker workspace, 10,800 s budget). Source: arxiv.org. Saturation forecast: Around 2029. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 20.68 | #75 |
| 2 | GLM-5.1 | 13.6 | #175 |
| 3 | DeepSeek V4 Pro | 13.19 | #181 |
| 4 | GPT-5.4 (xHigh) | 10.62 | #121 (GPT-5.4) |
| 5 | Gemini 3.1 Pro (Preview) | 10.12 | #91 |
| 6 | Kimi K2.6 | 9.7 | #173 |
| 7 | MiniMax-M2.7 | 9.26 | #519 |
| 8 | Qwen 3.6 Plus | 5.95 | #217 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=saasbench-engineering · How It Works · Data refreshed daily, snapshot 2026-11-12.