Rust code review: which models catch the bugs
Of the 17 models ranked on Rust code, Claude Opus 5.5, GPT-6.1 Sol and Claude Sonnet 5.5 tied for the most known bugs: 100%. 9 of the other 14 models have 95% intervals that overlap theirs. Gemini 3.5 Flash-Lite caught the fewest, 0%.
| Model | Caught here | Every bug | Difference |
|---|---|---|---|
| Claude Opus 5.5 | 100% | 90% | +10 pts |
| GPT-6.1 Sol | 100% | 89% | +11 pts |
| Claude Sonnet 5.5 | 100% | 77% | +23 pts |
| Gemini 3.1 Pro | 88%[63%–100%] | 67% | +21 pts |
| GLM-5.3 Flash (Workers AI) | 88%[63%–100%] | 61% | +27 pts |
| Gemini 4 Argon | 75%[25%–100%] | 82% | −7 pts |
| GPT-5.6 Sol | 75%[25%–100%] | 72% | +3 pts |
| Gemini 3.8 Flash | 75%[25%–100%] | 68% | +7 pts |
| GPT-6 Luna | 75%[25%–100%] | 67% | +8 pts |
| Claude Haiku 5.5 | 75%[25%–100%] | 64% | +11 pts |
| GLM-5.3 (Workers AI) | 73%[44%–100%] | 61% | +12 pts |
| GPT-OSS 120B (Workers AI) | 71%[25%–100%] | 40% | +31 pts |
| DeepSeek V4 Pro (Workers AI) | 54%[19%–88%] | 54% | 0 pts |
| DeepSeek V4 Flash (Workers AI) | 46%[13%–83%] | 49% | −3 pts |
| Kimi K2.7 Code (Workers AI) | 46%[17%–67%] | 38% | +8 pts |
| Gemma 4 26B A4B (Workers AI) | 38%[0%–75%] | 26% | +12 pts |
| Gemini 3.5 Flash-Lite | 0% | 11% | −11 pts |
On average these models caught 69% of known bugs in Rust code, more often than the 60% the same 17 caught across the whole benchmark: the third-easiest of the 12 languages.
GPT-OSS 120B (Workers AI) gains the most here: 71%, against 40% over every bug.
Gemini 3.5 Flash-Lite drops the most: 0%, against 11% over every bug.
The change “perf(libs/core): poll only ready handles and yield between I/O batches” in denoland/deno, a change of over 300 lines that landed in April 2026, introduced a concurrency bug that a later upstream commit fixed.
Claude Opus 5.5, GPT-6.1 Sol, Gemini 4 Argon, Claude Sonnet 5.5, GPT-5.6 Sol, Gemini 3.8 Flash, GPT-6 Luna, Claude Haiku 5.5, Gemini 3.1 Pro, GLM-5.3 Flash, GPT-OSS 120B (Workers AI) and Gemma 4 26B A4B caught it; GLM-5.3 and DeepSeek V4 Flash (Workers AI) caught it in part; DeepSeek V4 Pro (Workers AI), Kimi K2.7 Code (Workers AI) and Gemini 3.5 Flash-Lite missed it.