FDE-Bench - Greenfield: leaderboard
The 66 greenfield tasks of FDE-Bench (Ding et al., 2026), where the agent writes deployment configuration from application source and a specification alone. Agents share one four-tool scaffold; submitted Dockerfiles, manifests and monitoring configuration are rebuilt and redeployed in a clean sandbox and pass only if the build, readiness, behavior and specification-conformance check layers all pass, with no LLM judge. Resolved rate in percent; higher is better.
Metric: Resolved rate (%; the 66 greenfield tasks, authored from application source and specification alone; one episode per task, same four-tool scaffold (bash, write_file, read_file, submit), task-specific step, time and token caps; submitted artifacts are rebuilt and redeployed in a pristine sandbox and graded by four gated programmatic check layers (build, readiness, behavior, spec conformance); resolved only if all four pass). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.6 Flash | 62.1 | #72 |
| 2 | GPT-5.6 Terra | 59.1 | #60 |
| 3 | Claude Sonnet 5 (High) | 57.6 | #108 (Claude Sonnet 5) |
| 4 | GPT-5.6 Luna | 51.5 | #117 |
| 5 | DeepSeek V4 Pro | 50 | #181 |
| 6 | DeepSeek V4 Flash | 42.4 | #254 |
| 7 | Gemini 3.5 Flash Lite | 36.4 | #300 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=fde-bench-greenfield · How It Works · Data refreshed daily, snapshot 2026-11-12.