Phoenix-bench: leaderboard
Phoenix-bench (Zou et al., 2026) tests coding agents on 511 real hardware-engineering issues from 114 Verilog and SystemVerilog GitHub repositories: the agent edits the repository and the patch must pass fail-to-pass and pass-to-pass Verilator testbenches in a pinned Docker EDA environment. Resolved rate in percent, one attempt per issue without testbench feedback; each agent and backbone pair is its own row, higher is better.
Metric: Resolved rate (%; share of the 511 Phoenix-bench Verilog/SystemVerilog GitHub issues from 114 repositories whose agent patch passes the fail-to-pass and pass-to-pass Verilator testbenches in the pinned Docker EDA environment; one attempt per instance, no testbench feedback). Source: arxiv.org. Saturation forecast: Around 2028. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 38.6 | #75 |
| 2 | GPT-5.5 | 38 | #48 |
| 3 | Gemini 3.1 Pro (Preview) | 37.4 | #91 |
| 4 | GPT-5.2 | 33.9 | #172 |
| 5 | Gemini 3 Pro | 33.5 | #135 |
| 6 | GPT-5.3 Codex | 32.7 | #114 |
| 7 | Qwen 3 Coder 480B A35B Instruct | 32.3 | #641 |
| 8 | DeepSeek V3.2 | 26 | #403 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=phoenix-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.