FlinkEval v3.1 L0 - Pass and Ran: leaderboard
FlinkEval tests natural-language requests for Flink SQL. Ground-truth-anchored correctness, face-value judgments and execution errors are separate metrics. The main comparison uses 666 questions; the specification-ladder study uses a matched 600-question cohort, with each information condition retained. This board reports judge pass and successfully ran (%). Higher is better.
Metric: Judge pass and successfully ran (%). Source: arxiv.org. Saturation forecast: Around May 2027. 5 models tracked. Updated 14 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 68.3 | #20 |
| 2 | GPT-6 Astra | 68.3 | #6 |
| 3 | GPT-5.6 Sol | 59.9 | #30 |
| 4 | Gemini 3.6 Flash | 56 | #75 |
| 5 | GPT-OSS-120B | 34.8 | #703 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=flinkeval-v3-1-l0-pass-and-ran · How It Works · Data refreshed daily, snapshot 2026-11-16.