Google released Gemini 3.8 Live to developers on September 15, a real-time voice model that holds conversations in more than 97 languages for $0.005 per minute of audio input and $0.018 per minute of audio output. OpenAI charges $0.05 per session minute for GPT-Live-1, its rival model.
Key Takeaways
- Gemini 3.8 Live answers speech with speech and calls functions without breaking the conversation.
- An Extended Thinking variant tops Artificial Analysis’ speech-to-speech leaderboard, Google says.
- Even with both of Google’s rates added up, a minute costs less than half of an OpenAI session minute.
Have an AI Sum Up This Article
ChatGPTWhat Gemini 3.8 Live Does Without a Text Step
Google introduced three models in its post for developers building real-time voice applications. Gemini 3.8 Live is the centerpiece: a native speech-to-speech model that takes voice in and sends voice back, built for conversational agents.
The model fires function calls asynchronously, so the dialogue keeps going while a task runs in the background. It also reads visual context and aims for precision on alphanumeric strings, such as codes or technical references read out loud. Google says it covers more than 97 languages while keeping accents consistent.
The release lands about two weeks after Gemini 3.8 Flash came within three tenths of Claude Opus 5 on coding. The 3.8 generation now spans both code and spoken conversation.
A variant called Gemini 3.8 Live Extended Thinking adds configurable thinking for multi-step reasoning. Google says this version ranks first on Artificial Analysis’ speech-to-speech leaderboard.
The third model, Gemini 3.5 Transcribe, handles speech-to-text. Google lists more than 85 languages, a word error rate of 4.0% when streaming and 2.6% when not, automatic switching between languages mid-sentence, custom vocabulary of up to 1,000 terms, and a smart mode that formats the text and strips out filler words.
Gemini 3.8 Live Costs Less Than Half of GPT-Live-1
Google lists $0.005 per minute of audio input and $0.018 per minute of audio output for Gemini 3.8 Live. Those figures are estimates derived from token pricing of $3 per million input tokens and $12 per million output tokens.
On the other side, OpenAI’s GPT-Live-1 pricing page shows $0.05 per session minute, billed per second, with backend model and tool usage billed separately.
We ran the numbers in the scenario least favorable to Google, where audio flows nonstop in both directions for a full minute. Even then, Gemini 3.8 Live comes to $0.023, less than half of the $0.05 a GPT-Live-1 session minute costs. That math is ours, since Google does not publish a direct comparison.
The two pricing models do not measure the same thing. OpenAI bills session time, while Google bills audio volume in each direction. In a call where the agent listens more than it talks, most of Google’s bill falls on audio input, the cheapest line on the rate card.
Access runs through the Gemini API and Google AI Studio, with integrations at LiveKit, LangChain and Pipecat. Teams already building voice agents on those frameworks can try the model without rebuilding their pipeline. Google has also kept up a fast release pace, with Gemini 3.7 Flash replacing 3.6 after just three weeks this summer.
More articles on Horizon
- Siri AI on iPhone Leaves EU Users Waiting
- Jensen Huang Promises Trump No AI Slowdown
- Obama Wants Democrats to Get an AI Plan
Google Turns Voice Into a Price Fight
The launch fits a broader shift toward voice. Google Assistant shut down on September 4 across Android, watches and Android Auto, with Gemini taking over device by device.
Price is a familiar lever for Google. Gemini 3.6 Flash already shipped cheaper than the model it replaced, and the same playbook now reaches audio, where rates are quoted by the minute.
For OpenAI, the gap raises a blunt question. As of this writing, the GPT-Live-1 page still lists $0.05 per minute. A developer comparing both rate cards for a high-volume voice agent sees the difference in the very first budget estimate, and every extra minute of traffic widens it in absolute terms.
Distribution works in Google’s favor as well, since Gemini reported 950 million monthly users in the second quarter of 2026. The September 15 announcement targets developers through the API, though, and says nothing about the consumer app moving to Gemini 3.8 Live.
For teams handling calls at volume, the pairing with Gemini 3.5 Transcribe is part of the appeal. Custom vocabulary of up to 1,000 terms targets product names, acronyms and internal references specific to each company, and mid-sentence language switching covers conversations where a customer moves from one language to another.
Price does not settle everything. Perceived voice quality, real latency in production and how well function calls hold up over long conversations still need testing for Gemini 3.8 Live beyond the figures Google provides. Teams moving over from GPT-Live-1 will also have to rework their cost tracking, since a dashboard built on session minutes does not map directly onto audio minutes billed in each direction.
On budget, developers have a simple starting point: measure the split between incoming and outgoing audio on their own calls, since that split sets the real gap between the two rate cards. An agent that talks a lot moves toward the $0.023 ceiling, while one that mostly listens lands well below it. Running both models on the same sample of real calls for a week gives a better answer than any rate card comparison, and it also surfaces the quality and latency differences that pricing alone cannot show.
Follow the story on Horizon.
