OpenAI is preparing GPT-Bidi-1, a voice model that listens and speaks at the same time

The name Bidi is short for bidirectional. This is a fundamental difference from the current Advanced Voice Mode based on GPT-4o: that model was not created specifically for two-way real-time audio and works in a queue mode — first listens, then responds. GPT-Bidi-1 does both at the same time.

https://x.com/testingcatalog/status/2069331697615749530?s=20

Bidi can do several things at once:

• Speaks and listens at the same time — can continue answering while you are still talking;
• Handles interruptions and pauses naturally, without freezing and restarting;
• Switches between tasks in the middle of a sentence—for example, from counting forward to counting backward without stopping;
• Better maintains the context of a long conversation — one of the main weaknesses of the current voice stack;
• Supports three levels of «intelligence»: High, Medium, Instant — selectable in settings;
• The yellow color of the voice mode bubble instead of blue is a new visual identification.

According to testing data from TestingCatalog, the model produces short, natural confirmations — «okay,» a short «uh-huh» — without interrupting the other person. It also supports real-time translation, which opens up a whole class of new API-based products.

This is part of a broader strategy for OpenAI: text models have already reached the GPT-5.5 generation, while the voice stack has lagged behind. Bidi-1 is an attempt to close this gap and make voice a full interface, not an additional feature. The model is expected to appear in the Codex as well.

There is no official confirmation from OpenAI. It is also known that the current Advanced Voice Mode was not built for full-duplex audio, and Bidi-1 is an architectural leap, not just an update.

According to sources, it could happen literally this week. This is the answer to the question «how soon?».

Empero released an uncensored Qwythos-9B model — and said it trained it on data from the closed Claude Mythos

On the topic

Empero released an uncensored Qwythos-9B model — and said it trained it on data from the closed Claude Mythos

Read the country’s main IT news in our Telegram

On the topic

Read the country’s main IT news in our Telegram