# Google Ships Gemini 3.8 Live and Claims the Top Speech-to-Speech Score at 82.6

The two live dialogue models switch among 97 languages mid-conversation and run tool calls in the background, but agentic voice task completion stays below 70 percent on the benchmark Google cites.

- Published: 2026-09-16T06:18:31.385Z
- Canonical: https://polylog.news/ai/2026-09-16/google-ships-gemini-3-8-live-and-claims-the-top-speech-to-sp
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Google DeepMind](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/), [Polylog editors](https://polylog.news)

Google released [Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/) on 15 September, a pair of live dialogue models that reach developers through the Gemini application programming interface (API) and Google AI Studio, and reach end users through Search Live, Gemini Live, Gemini Enterprise and Google Workspace. The cheaper model is built for high-volume use. Extended Thinking is built for multi-step work: it reasons while still speaking, continuing tool calls and API requests in the background instead of pausing the conversation.

Google reports that Extended Thinking takes first place on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6, scores 97.7 percent on Big Bench Audio, and leads agentic voice task completion at 68.6 percent on τ-Voice and 35.1 percent on Sierra's τ-Voice-banking. The standard model placed second in Artificial Analysis' Speech Agent Arena. Both models detect and switch among 97 languages mid-conversation and accept camera input in near real time, as [summarized in Russian-language coverage](https://t.me/ai_machinelearning_big_data/10935) and [in MarkTechPost's write-up](https://www.marktechpost.com/2026/09/15/google-releases-gemini-3-8-live-and-3-8-live-extended-thinking-for-production-grade-voice-agents/).

The agentic-task numbers are more informative than the quality index. A 35.1 percent completion rate on banking voice tasks means the best-scoring model fails roughly two of every three scripted financial interactions end to end, which is the gap between a convincing demo and an unsupervised customer-service deployment. The index placements come from an independent evaluator, but the selection of which benchmarks to publish is Google's, and no external reproduction of the full set exists yet. Google also describes the pricing as competitive with rival frontier models without committing to a published comparison.

## What this means

Voice is becoming the product category where Google converts its distribution advantage into a share of the API market, because Search Live, Workspace and Gemini Enterprise give it live-audio traffic that OpenAI and Anthropic have to buy instead. The mechanism for rivals is cost: a model that holds a conversation while executing tool calls in the background removes the latency padding that voice vendors currently sell as orchestration, which reduces the pricing power of the middleware layer built on top of speech APIs. The immediate commercial test is whether contact-center buyers accept the τ-Voice-banking failure rate with human escalation, or whether they wait for completion rates that support unattended operation.

## What to watch

- Independent reproduction of the 82.6 speech-to-speech score and the τ-Voice agentic results by evaluators Google did not cite, which would separate a real capability step from selective benchmark disclosure.
- Published per-minute or per-token pricing for both models, since Google's cost claim is currently an assertion and pricing is where this release pressures OpenAI's and Amazon's voice offerings.
- Whether regulated buyers such as banks and insurers move voice agents from pilots to unsupervised handling, which is the signal that live agentic models have reached the reliability threshold that matters commercially.