Gemini 3.8 Live Hits $0.023/Min: Google Undercuts Voice AI Stack

Google shipped Gemini 3.8 Live on September 15 at $0.005 per minute for audio input and $0.018 per minute for output. That is $0.023 per minute of two-way conversation, or roughly $1.38 per hour of live voice agent time, before you add tool-call latency, streaming infrastructure, or telephony egress. For anyone running a cascaded speech-to-text plus LLM plus text-to-speech pipeline today, that number is the one to benchmark against this quarter.

What Happened

As blog.google reported, Google DeepMind released three new audio models into the Gemini API and Google AI Studio: Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking, and (shipped a month earlier) Gemini 3.5 Transcribe. The post was authored by Alisa Fortin, a Product Manager at DeepMind, and Thor Schaeff from the DevX technical staff.

The headline capabilities: 3.8 Live is a native speech-to-speech model that can execute tasks while holding a dialogue. The Extended Thinking variant adds configurable background reasoning for multi-step requests and, per Google, currently ranks #1 on Artificial Analysis' Speech-to-Speech leaderboard. Language coverage is 97+ for the Live models and 85+ for Transcribe. 3.5 Transcribe posts a 4.0% average Word Error Rate in streaming mode and 2.6% in non-streaming, with custom vocabulary biasing up to 1,000 terms and a Smart mode that strips disfluencies.

Distribution is deliberate. The Live models ship behind the Live API with named integration partners handling the media plane: Agora, Fishjam, LiveKit, LangChain, Pipecat, Vercel, and Vision Agents. Transcribe is also reachable through the Interactions API, which accepts audio files up to one hour with structured timestamps and speaker labels. Google rounded out the announcement by grouping the launch with three adjacent models already in the API: Gemini 3.5 Live Translate (70+ languages, speech-to-speech), Gemini 3.1 Flash TTS, and Lyria 3.5 for music generation.

Technical Anatomy

The important architectural claim is that 3.8 Live is native speech-to-speech, not a cascade. A traditional voice agent stack looks like this: audio in, ASR model, LLM, TTS model, audio out. Each hop adds latency (typically 150 to 400 ms per stage in production) and each hop drops information: prosody, hesitation, code-switching cues, background sound. A native speech-to-speech model collapses those stages into a single forward pass, which is why the Extended Thinking variant can "narrate its progress" while reasoning in the background rather than blocking on a text turn.

Five capabilities matter for engineering teams evaluating this. Asynchronous function calling lets the model keep streaming audio to the user while a tool call runs in the background, which is the single largest UX defect in current voice agents (the awkward silence while the LLM waits on a database query). Visual context grounding accepts live visual input alongside audio. Alphanumeric precision is called out explicitly for confirmation codes and claim numbers, a nod to the vertical Google clearly wants: contact centers. Incremental content updates merge structured data into the audio stream mid-response. Configurable thinking on the Extended Thinking variant is the lever developers will use to trade latency for reasoning depth.

On the transcription side, the 2.6% non-streaming WER is the number that resets industry expectations. For comparison, human transcribers typically score in the 4 to 5% range on conversational English. The 4.0% streaming figure matters more for real-time captioning and call center analytics, since streaming WER is normally 1 to 2 percentage points worse than batch. Custom vocabulary biasing with 1,000 terms is a meaningful upper bound, though Google does not disclose whether biasing degrades general recognition accuracy when the list is heavily loaded, which matters because that is the failure mode teams actually hit in production. We do not know yet how sensitive the WER is to noisy telephony audio (8 kHz narrowband), and the source does not break WER down by language across the 85+ supported.

Who Gets Burned

The most exposed category is the standalone real-time transcription vendors whose entire moat is streaming ASR accuracy. When a hyperscaler ships 4.0% streaming WER as a general-purpose API, the pricing pressure on specialized ASR providers is one-directional. Expect their sales conversations over the next two quarters to shift from accuracy benchmarks to compliance, on-prem deployment, and industry-specific vocabularies where the 1,000-term bias list runs out.

Second: builders of cascaded voice agent frameworks. The named Live API partners (LiveKit, Pipecat, Vercel, Agora, Fishjam, LangChain, Vision Agents) are positioned as media plane and orchestration layers, not as model providers. That is a survivable position. Framework projects that tried to be opinionated about the ASR-to-LLM-to-TTS orchestration itself now have a competitor whose latency floor they cannot match, because they still pay for three network hops where 3.8 Live pays for one.

Third: iGaming and fintech voice-verification workflows. Alphanumeric precision on confirmation codes and claim numbers is Google aiming directly at KYC callbacks, deposit confirmations, and support triage. The 97+ language coverage matters here because these are exactly the workflows where a European-facing operator serves customers in fifteen locales and does not want fifteen separate ASR contracts.

Fourth: OpenAI's Realtime API is now competing against a published leaderboard rank and a specific per-minute price. The OpenAI docs list Realtime pricing separately for audio input and output tokens, which makes direct comparison awkward for procurement teams. Google's flat per-minute pricing is easier to model in a spreadsheet, and that alone will influence RFP decisions. My prediction: within 90 days we should see at least one major voice agent vendor publish a public bakeoff showing 3.8 Live latency versus their existing pipeline. If they do not, assume the numbers are unflattering.

Playbook for AI Development

Concrete actions for engineering leads this week. First, run a cost model. Take your current voice agent's average session length and multiply by $0.023 per minute of two-way audio. If your existing STT plus LLM plus TTS bill per session is more than roughly 1.5x that number, you have a migration business case worth writing down. If it is under, the switching cost probably is not justified yet.

Second, benchmark WER on your actual audio, not on LibriSpeech or the vendor's marketing clips. The 2.6% non-streaming figure is an average across Google's evaluation set. Your call center audio with three overlapping speakers on a noisy line will not hit that. Pull a 200-sample corpus from production, run it through 3.5 Transcribe via the Interactions API, and compare to your incumbent.

Third, test asynchronous function calling against a realistic tool graph. The demo case is easy; the production case is a customer asking a question that requires two sequential database lookups and a third-party API call. Measure how the model narrates progress and whether it hallucinates while the tool call is pending.

Fourth, if you are building agent orchestration, treat the Live API partner list as a signal about where Google expects the media plane to live. Integrating with one of those seven partners is now the lower-risk architectural choice than rolling your own WebRTC layer. Look at MCP patterns for the tool-call side so you are not locked to one model vendor's function-calling schema.

Key Takeaways

• Gemini 3.8 Live prices at $0.023/min of two-way audio ($0.005 input, $0.018 output), setting a public benchmark for cascaded voice stack economics.
• Gemini 3.5 Transcribe posts 2.6% non-streaming WER and 4.0% streaming WER across 85+ languages, pressuring standalone real-time ASR vendors.
• Native speech-to-speech with asynchronous function calling addresses the awkward-silence problem that plagues cascaded voice agents during tool calls.
• Live API launches with seven named media-plane partners including LiveKit, Pipecat, and Agora, signaling Google's preferred deployment topology.
• Unknowns worth testing: WER degradation on 8 kHz telephony audio, accuracy trade-off when custom_vocabulary is loaded near the 1,000-term ceiling, and per-language WER variance across the 85+ supported.