Gemini 3.8 Live Extended Thinking

Why teams choose it

Tradeoffs to know

• Same Live API gaps as 3.8 Live: no caching, code execution, structured outputs, or image generation on the model card.
• Vendor Speech-to-Speech and τ-Voice scores are Google-cited; reproduce on your own live harness before replacing 3.8 Live.

When not to use this

• Do not use this as a drop-in for Gemini 3.8 Flash coding routes or for ASR-only jobs (use Gemini 3.5 Transcribe).
• Not ideal for simple tasks where cheaper models in the same lineup are good enough.
• Avoid for latency-sensitive real-time chat when raw response speed outweighs reasoning depth.

Technical specs

Inputs

text, image, audio, video

Outputs

audio, text

Capabilities

full-duplex voice, background reasoning, asynchronous function calling, progress narration, visual grounding, search grounding, audio generation, configurable thinking

License

Proprietary API

Model string

gemini-3-8-live-extended-thinking

Benchmarks

{
  "api_id": "gemini-3.8-live-extended-thinking",
  "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/",
  "tau_voice": 68.6,
  "context_tokens": 131072,
  "big_bench_audio": 97.7,
  "vendor_reported": true,
  "max_output_tokens": 65536,
  "tau_voice_banking": 35.1,
  "audio_input_usd_per_minute": 0.005,
  "audio_output_usd_per_minute": 0.018,
  "aa_speech_to_speech_quality_index": 82.6
}