Grok Voice Think Fast 2.0 beats GPT Realtime on VulcanBench, 99% text accuracy
XFreeze · x · 2026-08-20
Per X user XFreeze, Grok Voice Think Fast 2.0 beat GPT Realtime across both text and voice on VulcanBench: 99% accuracy on 200 held-out text questions, with stronger performance too when the same questions were spoken.
The voice result stands out — Grok preserves nearly all of its performance moving from typed to spoken input while still completing full responses dramatically faster. The author argues Grok is becoming a seriously powerful realtime reasoning model that can listen, think and respond incredibly fast. (Unofficial benchmark, unverified.)
More from Models
- Grok Bot Ran a 12-Hour Data Analysis Nonstop, Handling API Rate Limits Itself — Daniel_Farinax · 2026-10-16
- One year of Recursive Language Models: GPT-5-mini beats GPT-5 by 110% on 132k-token tasks, cheaper too — a1zhang · 2026-10-16
- Users Report Claude Models Suddenly More Compliant After Forced Compaction — RealSharpNinja · 2026-10-16
- Prediction: Opus 5.5-level models on laptops by 2027, backed by Qwen 27B data — cephaloform · 2026-10-16
- Users Say Krea 2 Output Degraded Sharply Since Krea 3 Beta Launched — szywis · 2026-10-16
- LLM-as-a-Verifier lifts Nemotron 3.5 Super by 9-10 points on agentic benchmarks — simonguozirui · 2026-10-16