Artificial Analysis: Gemini 3.8 Live tops voice agent benchmarks at record low cost
On 09-16, Artificial Analysis published multi-dimension voice benchmarks for Gemini 3.8 Live, showing leading or near-leading performance across latency, agent capability, audio reasoning, and cost. Combined with Google's touted "talk while working" capabilities, the results point to intensifying competition in real-time voice agents.
Confirmed
- Latency: Gemini 3.8 Live averages 1.18s first-audio latency on Big Bench Audio (1.35s for Extended Thinking High), roughly 2.5x faster than the previous generation.
- Agent capability: Gemini 3.8 Live Extended Thinking (High) tops the Tau Voice agent benchmark at 68.6%, nearly double its predecessor and ahead of GPT-Live.
- Audio reasoning: 97.7% on Big Bench Audio, ahead of Grok Voice Think Fast and second only to Qwen.
- Human preference: debuted at Elo 1083 in the Speech Agent Arena blind test, ranking 2nd behind the previous-gen Gemini 3.1 Flash Live.
- Pricing: input audio at $0.84/hour is the cheapest in its Speech to Speech Index, roughly half the price of Gemini 3.1 Flash Live High.
- Per @heyshrutimishra, Extended Thinking supports thinking, tool calls, and background task completion while speaking, continuously syncing progress by voice.
Why it matters
- The third-party data spans five dimensions—latency, agent benchmarks, audio reasoning, blind preference, and cost—providing a comprehensive comparison profile.
- The combination of low latency and low cost lowers the barrier for real-time voice agent applications, likely intensifying competition with GPT-Live, Grok, and other voice models.
2026-09-16 ~ 2026-09-17 · 7 related posts
Primary sources
- Gemini 3.8 Live hits 1.18s Time to First Audio, 2.5x faster than its predecessor — ArtificialAnlys ·
- Gemini 3.8 Live tops Tau Voice agentic benchmark at 68.6%, up from 37.7% — ArtificialAnlys ·
- Gemini 3.8 Live Can Talk While It Works: Tops Speech-to-Speech Index at 82.6 — heyshrutimishra ·
- Gemini 3.8 Live debuts #2 in Speech Agent Arena with Elo 1083, 93.2% task success — ArtificialAnlys · 2026-09-16
- [source] Gemini 3.8 Live tops Tau Voice agentic benchmark at 68.6%, up from 37.7% — ArtificialAnlys · 2026-09-16
- Gemini 3.8 Live scores 97.7% on Big Bench Audio, second only to Qwen — ArtificialAnlys · 2026-09-16
- [source] Gemini 3.8 Live hits 1.18s Time to First Audio, 2.5x faster than its predecessor — ArtificialAnlys · 2026-09-16
- Gemini 3.8 Live priced at $0.84/hour, cheapest speech model on the index — ArtificialAnlys · 2026-09-16
- [source] Gemini 3.8 Live Can Talk While It Works: Tops Speech-to-Speech Index at 82.6 — heyshrutimishra · 2026-09-17
1 near-duplicate retellings: ArtificialAnlys