---
title: "Realtime-Bench reveals low-latency limits of current LLMs"
slug: "realtime-bench-reveals-low-latency-limits-of-current-llms"
date: "2026-11-04"
category: "AI & Automation"
author: "Dr. Kai Wei"
correspondent: "Dr. Kai Wei"
correspondentRole: "Lead Frontier AI Correspondent"
readingTimeMinutes: 2
tags: ["AI","News"]
originalSourceUrl: "https://importai.substack.com/p/import-ai-479-spacewar-low-latency"
canonicalUrl: "https://aurav.ae/realtime-bench-reveals-low-latency-limits-of-current-llms"
publisher: "Aura-V Intelligence (aurav.ae)"
---

# Realtime-Bench reveals low-latency limits of current LLMs

> Researchers introduced Realtime-Bench, a 50-task suite measuring how well leading foundation models handle real-time decisions, exposing sub-70% success on timed challenges.

## Key Takeaways (TL;DR)

- Realtime-Bench tests 50 interactive tasks to gauge temporal awareness, adaptive deliberation, and time-aware action selection in LLMs.
- GPT-6 Astra tops the wall-clock condition with a 36% pass rate, far below paused-clock performance of 70-80%.
- The benchmark highlights that current models struggle with real-time applications such as ticket buying or online auctions.

Researchers from the University of Texas at Austin, Ohio State, UIUC, NVIDIA, Northwestern and Steadyworks released Realtime-Bench, a suite of 50 dynamic tasks—30 web-based (travel, e-commerce, service, security) and 20 games (arcade, puzzle, strategy, simulation)—to probe AI agents’ low-latency capabilities. The benchmark measures temporal state awareness (RT-Perceive), adaptive deliberation (RT-Deliberate) and time-aware action selection (RT-Act) across three clock settings: wall clock, paused clock and token clock. The design forces models to track changing environments, limit reasoning as deadlines approach, and time actions appropriately, as illustrated by a ticket-purchase scenario.

The authors evaluated a range of models, including OpenAI’s GPT-6-Astra, GPT-5.6 Sol, GPT-Realtime-2, OpenAI Decisions, Anthropic’s Claude Fable 5.1 and Claude Opus 5, Google’s Gemini 3.8 Flash, Gemini 3.1 Flash Live, Gemini 3.8 Live, Jev 1.13, Deepseek V4.1 and Moonshot AI’s Kimi-K3. GPT-6 Astra achieved the highest wall-clock pass rate at 36%, while paused-clock scores rose to 70-80% for high-reasoning models, underscoring the latency gap.

Our read: The gap between paused-clock and wall-clock performance suggests that inference speed, not just model size, remains a critical bottleneck for deploying LLMs in real-time B2B workflows.

---

*Original Think-Tank Source: [Import AI 479: Spacewar; low-latency tasks versus LLMs; Genie 4](https://importai.substack.com/p/import-ai-479-spacewar-low-latency)*