Gemini 3.8 Live vs Extended Thinking vs GPT-Live-1 Compared
The short answer
| Gemini 3.8 Live | Gemini 3.8 Live Extended Thinking | GPT-Live-1 (OpenAI) | |
|---|---|---|---|
| Released | September 15, 2026 | September 15, 2026 | ChatGPT July 2026 · API September 10, 2026 |
| Architecture | Native speech-to-speech, single model | Native speech-to-speech + configurable background thinking | Full-duplex voice layer + separate backend text model |
| Audio price | $0.005/min in · $0.018/min out | $0.005/min in · $0.018/min out | $0.05/min voice layer + backend model tokens |
| AA Speech to Speech Index | Not listed at launch (2nd in Speech Agent Arena) | 82.6 (#1) | 81.5 with GPT-6 Astra (medium) · 80.1 with GPT-5.6 Sol (low) |
| τ-Voice (agentic) | — | 68.6% · 35.1% τ-Voice-banking | Published by AA per backend config |
| Big Bench Audio | — | 97.7% | — |
| Languages | 97, auto-switching mid-call | 97, auto-switching mid-call | Broad, not enumerated |
| Vision in the call | Yes, near real-time | Yes | No native visual grounding |
| Interruptions | Turn-aware; background tool calls keep the conversation going | Same, plus “Let me check that…” narration | Native full-duplex (listens while speaking) |
| Open weights | No | No | No |
| Best for | High-volume, cost-sensitive agents; visual context | Complex multi-step voice workflows (banking, support, bookings) | Interruption-heavy conversations; picking your own reasoning model |
Rule of thumb, as of September 17, 2026: Extended Thinking for complex agentic voice work, plain 3.8 Live for volume, GPT-Live-1 when turn-taking realism is the product.
What Google shipped on September 15, 2026
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026, describing them as its “most advanced live dialogue models yet.” Both are native speech-to-speech models — one network hears audio and produces audio — rather than a cascade of speech-to-text, LLM and text-to-speech. Google positions them explicitly against cascaded pipelines, which add latency at every handoff and lose prosody.
The two models split by job:
- Gemini 3.8 Live is “built for scale and cost efficiency.” It combines dialogue quality with near-real-time visual grounding (the model can look at a camera feed or screen share while it talks), executes tool and API calls in the background so it can acknowledge a request and keep chatting, and auto-detects and switches between 97 languages mid-conversation with accent consistency.
- Gemini 3.8 Live Extended Thinking is “built for high-complexity tasks.” It reasons and speaks simultaneously: it uses early verbal cues (“Let me check that…”), narrates progress on multi-step background tasks, and exposes configurable thinking in the Live API. Google’s demos show it turning a sketch plus spoken feedback into working React components and coordinating multi-step bookings.
Both models ship through the Gemini Live API and Google AI Studio for developers, in private preview in Gemini Enterprise, and via Live API partners — Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel AI Gateway and Vision Agents. Salesforce, Genspark and Lumeris are named launch partners. All generated audio carries a SynthID watermark.
Benchmarks: Extended Thinking takes the top spot
Google’s headline claim is that Gemini 3.8 Live Extended Thinking scores 82.6 on Artificial Analysis’ Speech to Speech Index, the #1 overall position. The index is a weighted average of four components: speech reasoning (Big Bench Audio), agentic performance (τ-Voice), human arena preference and task success rate.
The component scores Google published:
- τ-Voice: 68.6% — the share of replica customer-service scenarios resolved while acting as the support agent
- Sierra’s τ-Voice-banking: 35.1% — a harder, domain-specific variant
- Big Bench Audio: 97.7% — speech reasoning
Before this release, GPT-Live-1 paired with GPT-6 Astra (medium reasoning) held #1 at 81.5, with the GPT-5.6 Sol (low) configuration at 80.1. The margin is about a point, which is within the range where your backend model choice on the OpenAI side matters as much as the voice layer. Plain Gemini 3.8 Live placed second in the Speech Agent Arena, a blind human-preference test where participants complete the same task (booking a dental appointment, say) with two unidentified models.
On ServiceNow’s EVA-Bench Google says both models push the Pareto frontier for complex workflows — balancing task accuracy against conversational quality — with the caveat that the run used the Live API on Gemini Enterprise Agent Platform, not the public endpoint.
Pricing: Gemini undercuts GPT-Live-1 on the voice layer
Google lists both 3.8 Live models at $0.005 per minute of audio input and $0.018 per minute of audio output, an estimate derived from $3 per million input tokens and $12 per million output tokens. A support call in which the agent speaks roughly half the time therefore lands around $0.011–$0.012 per minute, and a monologue-heavy tutoring session tops out near $0.018.
GPT-Live-1 in the OpenAI API (September 10, 2026) charges $0.05 per minute for the voice layer, billed per second — and that excludes the backend text model that does the reasoning and tool calls. With GPT-5.6 Luna ($0.20/$1.20 per MTok) the backend bill is small; with GPT-6 Astra ($10/$50 per MTok) it can rival the voice charge on a long, tool-heavy call. OpenAI’s turn-based alternative, GPT-Realtime-2.1, bills $32/$64 per million audio tokens (roughly $0.05/min), and Realtime-2.1-mini about $0.016/min.
So at list price, Gemini 3.8 Live is roughly 3–5x cheaper per conversation minute than GPT-Live-1 before OpenAI’s backend tokens are counted. The usual caveat applies: real spend depends on how much the model talks, and Extended Thinking’s background reasoning consumes output tokens Google has not itemised separately.
Where GPT-Live-1 still wins
- True full-duplex. GPT-Live-1 is a single model over both audio streams, listening while it speaks. OpenAI reports a 30-percentage-point lead over GPT-Realtime-2.1 on Full Duplex Bench. Gemini 3.8 Live handles interruptions well but is architected around background tool calls rather than simultaneous listen-and-speak.
- Bring-your-own brain. GPT-Live-1 delegates reasoning to a backend model you choose — GPT-6 Astra, GPT-5.6 Terra or Luna, or a third-party model. Gemini’s reasoning is the Gemini model; there is no swap.
- Telephony maturity. GPT-Live-1 ships with telephony, ASR transcripts, keyword biasing and explicit turn detection out of the box.
Where Gemini 3.8 Live wins
- Visual grounding in the call. Neither GPT-Live-1 nor Realtime-2.1 offers Gemini’s near-real-time processing of camera or screen input during a conversation.
- 97-language auto-switching. A caller who drifts from English into Spanish mid-sentence does not trigger a reconfiguration.
- Alphanumeric precision. Google specifically calls out reliable parsing of confirmation codes, claim numbers and technical strings — the classic failure point in voice IVR.
- Price and consumer reach. Same-day availability in Search Live, Gemini Live, Gmail, Keep and Docs means the models are exercised at consumer scale from day one.
Sub-questions people ask
Can I self-host either model? No. Both Gemini 3.8 Live models and GPT-Live-1 are hosted-only.
Does Extended Thinking add latency? Google’s design goal is that it does not block speech — the model acknowledges immediately and narrates while it reasons. Expect the task to take longer than plain 3.8 Live on complex requests, not the first audio.
Is Gemini 3.8 Live the same as Gemini 3.8 Flash? No. Gemini 3.8 Flash (September 2, 2026) is the text/multimodal model at $0.75/$3.75 per MTok introductory pricing; the Live models are the audio family, which also includes Gemini 3.5 Transcribe.
Which one for Google Workspace users? Extended Thinking is rolling out to Google AI Pro and Ultra subscribers in Docs and to all Google AI subscribers in Gmail and Keep; Workspace business customers are “coming soon” via Gemini Enterprise.