BackendForge (Base Oracle): leaderboard
BackendForge (Peking University, 2026) asks coding agents to build a complete Dockerized Python backend service from a written specification and an OpenAPI contract for 56 services rewritten from open-source applications, judged only through black-box HTTP tests. Task success rate in percent against the base oracle written before any reference implementation existed; a task counts only when every test passes, higher is better.
Metric: Task success rate (%; base oracle of 7,250 items written before the reference implementation; 56 contract-defined backend services rewritten from open-source applications; from a visible spec and OpenAPI contract the agent builds a Dockerized Python service, scored only through black-box HTTP pytest items; a task counts only when every item passes; one attempt per task in the same mini-SWE-style coding harness). Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (Max) | 58.93 | #79 (Claude Opus 4.7) |
| 2 | GPT-5.5 (xHigh) | 55.36 | #49 (GPT-5.5) |
| 3 | Kimi K2.6 | 39.29 | #178 |
| 4 | Claude Sonnet 4.6 (Max) | 39.29 | #161 (Claude Sonnet 4.6) |
| 5 | DeepSeek V4 Pro (Max) | 32.14 | #188 (DeepSeek V4 Pro) |
| 6 | GLM-5.1 | 26.79 | #179 |
| 7 | DeepSeek V4 Pro (High) | 25 | #188 (DeepSeek V4 Pro) |
| 8 | Qwen 3.6 Plus (High) | 23.21 | #222 (Qwen 3.6 Plus) |
| 9 | Gemini 3.5 Flash | 21.43 | #101 |
| 10 | DeepSeek V4 Flash (High) | 21.43 | #258 (DeepSeek V4 Flash) |
| 11 | DeepSeek V4 Flash (Max) | 19.64 | #258 (DeepSeek V4 Flash) |
| 12 | Hy3-preview (High) | 12.5 | #316 (Hy3-preview) |
| 13 | MiniMax-M2.7 | 5.36 | #518 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=backendforge-base-oracle · How It Works · Data refreshed daily, snapshot 2026-11-13.