SEABED - Long-Form Reasoning (Open-Ended): leaderboard
SEABED (Rajgarhia et al., 2026) is an audio-first reasoning benchmark on Southeast Asian speech built from openly available corpora. In this task the model answers questions over whole 7 to 23 minute Thai, Indonesian and Malay conversations that combine evidence from distant parts; answers are open-ended and a Grok 4.5 judge grades answers from 0 to 1, and the score is accuracy in percent, higher is better.
Metric: Accuracy (%; 182 reasoning over whole 7 to 23 minute Thai, Indonesian and Malay conversations items; open-ended answers judged by Grok 4.5, binary correct or incorrect (graded 0 to 1 for long-form); audio at 16 kHz). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.6 Flash | 68.3 | #72 |
| 2 | Gemini 3.5 Flash | 66.7 | #98 |
| 3 | Gemini 2.5 Pro | 64.6 | #262 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=seabed-long-form-reasoning-open-ended · How It Works · Data refreshed daily, snapshot 2026-11-12.