0
Neue Alchemy
Sign inSign upLet’s talk
Signal + NoiseIntelligence, distilled daily.
MODEL SIGNAL · GOOGLE DEEPMIND · NEW

Gemini 3.8 Live Extended Thinking

A natively multimodal Gemini 3.8 Audio model variant from Google DeepMind, optimized for real-time dialogue, that adds high‑reasoning background thinking for complex, multi‑step problem solving during live voice interactions.

CATEGORYMultimodal
CONTEXT131072 (input token limit), 65536 (output token limit)
RELEASEDSeptember 15, 2026
Key Features
  • Natively multimodal Gemini 3.8 Audio variant (inputs: text, images, audio, video; outputs: text and audio)
  • High‑reasoning background thinking with configurable thinking level (low, medium, high) for complex, multi‑step problem solving in real‑time voice sessions
  • Live API support for continuous streamed audio responses while performing background reasoning and asynchronous tool calls
  • Input token limit of 131,072 tokens and output token limit of 65,536 tokens
  • Optimized for real-time dialogue and high‑complexity tasks, including multi‑step diagnostics, complex planning, and multi‑tool workflows

Provider announcement →

Read the Model Signal report →

MODEL SIGNAL

Gemini 3.8 Live Extended Thinking

Google DeepMind bridges complex background reasoning with real-time multimodal audio streams.

Bottom line

Google DeepMind has introduced Gemini 3.8 Live Extended Thinking, a natively multimodal variant of its 3.8 Audio model optimized for real-time dialogue. The release brings configurable, high-level background reasoning directly to live voice sessions. The model supports text, image, audio, and video inputs, and generates continuous text and audio outputs via the Live API. It operates with a 131,072-token input limit and a 65,536-token output limit, allowing it to perform asynchronous tool calls and multi-step problem solving without breaking conversational flow.

Signal

The clear operator read here is the successful marriage of extended reasoning pipelines with streaming dialogue architectures. Historically, injecting deep, multi-step problem solving into a voice agent resulted in dead air or disrupted conversational rhythms. By introducing a configurable thinking level (low, medium, high) that interfaces natively with the Live API, Google is signaling that developers no longer have to choose between complex asynchronous tool execution and real-time user engagement. The model is structurally designed to maintain continuous streamed audio responses while processing complex logic in the background.

Noise

The noise lies in conflating this model with traditional text-based reasoning engines. Because it shares the "Extended Thinking" nomenclature with heavier, offline-style text models, operators might mistakenly prioritize standard static logic benchmarks. Its true differentiation requires leveraging the live audio loop and multimodal inputs. Furthermore, treating the "high" thinking configuration as a universal default will likely introduce unnecessary latency into simpler interactions; the value depends on dynamically matching the thinking configuration to the task's actual complexity.

Where it fits

This model is built for high-complexity, real-time voice orchestration. It fits optimally into multi-step technical diagnostics, interactive complex planning, or dynamic environments where a user talks through a problem while the agent queries databases, reasons over live visual or video inputs, and verbally synthesizes its findings. It is fundamentally over-engineered for basic conversational voice bots, standard customer service routing, or text-only reasoning tasks.

Model Signal · Signal + Noise · Isaiah Steinfeld