Google made Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking available through Gemini API.
The move continues AI systems' shift from text-based chat toward interfaces where users can speak to the system, receive voice responses and perform tasks in real time.
Both models are intended for real-time interaction, especially voice uses.
The detail making the launch particularly interesting to Israeli audiences is that Google explicitly lists Hebrew among Live API's supported languages.
The two models
Google presents two models:
Gemini 3.8 Live
Gemini 3.8 Live Extended Thinking
The standard model is intended for real-time conversations where a quick response matters.
Extended Thinking is intended for situations where the system needs to do more thinking during the conversation.
Both models can receive:
- Text
- Audio
- Images
- Video
And they can return:
- Text
- Audio
Hebrew appears on the official list
One problem in AI tool coverage is that "multilingual" sometimes automatically becomes "supports Hebrew".
In this case, there is a stronger basis.
Google explicitly lists Hebrew with the language code he among Live API's supported languages.
This still does not mean Hebrew quality is perfect.
Technical support is not the same as speech quality, accent, pauses, slang or understanding natural Hebrew.
But at least here Google documents Hebrew as supported rather than leaving it to inference.
How much it costs
According to the Gemini Developer API price list, as of September 21, 2026:
- Text input: 0.75 dollars per million tokens
- Text output: 4.50 dollars per million tokens
- Audio input: 3 dollars per million tokens or 0.005 dollars per minute
- Audio output: 12 dollars per million tokens or 0.018 dollars per minute
- Image or video input: 1 dollar per million tokens or 0.002 dollars per minute
Google also presents a free tier.
A conversation's actual price may depend on the mix of input and output types and its length.
Do not use only the price per voice minute to automatically estimate the entire session cost.
Why this matters for voice products
Voice is gradually turning from a supplementary interface into a real working interface with AI.
The challenge is no longer just getting a model to recognize speech.
A useful system needs to understand context, respond quickly, use tools and successfully complete tasks during a conversation.
Some of the most interesting AI agent uses are moving to voice.
For example:
A service center
A sales agent
A private tutor
A support system
A voice assistant for employees
An accessibility interface
An agent guiding users inside an app
For an Israeli company, official Hebrew support removes at least one barrier to experimentation.
What about ordinary users in the Gemini app?
The announcement concerns Gemini API.
Do not infer that every Gemini app user received the same models, capabilities and terms on the same day.
Distinguish between:
Gemini API for developers
And:
The Gemini app for consumers
What to test in Hebrew
Developers building a voice product can begin testing both models through the API.
For an Israeli product, conduct a real Hebrew test including:
- Fast speech
- Interruptions mid-sentence
- Multiple speakers
- Noise
- Israeli names
- Mixing Hebrew and English
- Numbers
- Addresses
- Professional terminology
Official support is a start.
Actual quality needs testing.
What has not yet been tested
At publication, no broad independent test of Gemini 3.8 Live's Hebrew quality was found.
Regional data processing options for Israeli companies were not verified either.
The effective cost of Extended Thinking in long conversations also depends on usage, not just the base price.