Recent AI news · Original report: 2026-09-16 (KST)

OpenAI GPT-Live-1 and Google Gemini 3.8 Live: real-time voice AI launches (Sep 10-15, 2026)

This article was compiled with AI. This English translation preserves the original report's collection date, findings, and limitations. Please consult the linked sources for current information. Originally collected: 2026-09-16 10:25 (KST, UTC+09:00).

Translated into English: 2026-09-28 (KST). The translation date does not indicate a new source check.

At a glance

On September 10 and 15, OpenAI and Google respectively released voice models for real-time conversation. In the same week, Anthropic published a report on misuse of its models, including a case in which someone’s voice was cloned and used in live conversations.

This week’s announcements

The organizations’ official announcements were opened and reviewed on September 16, 2026, at 10:25 a.m. KST.

OpenAI: a model that listens while speaking

Voice assistants have traditionally joined three stages: converting speech to text, reasoning, and converting text back to speech. OpenAI’s announcement says latency builds up at each stage and this setup can lose the flow of conversation. It describes GPT-Live-1 as handling listening and speaking together in one model.

• It lists a price of $0.05 per minute for the front-end voice layer. The model doing the reasoning behind it costs extra.
• It reports an improvement of 30 percentage points over the previous model, GPT-Realtime-2.1, on its internal Full Duplex Bench evaluation. The announcement does not give absolute scores.
• It cites an early evaluation by language-learning service Speak, in which interruptions while learners were thinking fell by nearly 80% compared with the previous approach.
• It says the model can be used for voice agents handling phone calls.

Google: continuing the conversation while work runs in the background

Google DeepMind released Gemini 3.8 Live and 3.8 Live Extended Thinking on September 15. All the benchmark scores quoted in its announcement are for the Extended Thinking version.

• It reports first place with a score of 82.6 on Artificial Analysis’s Speech to Speech Quality Index.
• It lists 68.6% on τ-Voice, 35.1% on Sierra’s τ-Voice-banking, and 97.7% on Big Bench Audio.
• The description of Gemini 3.8 Live says it automatically recognizes and switches between 97 languages during conversation.
• It says tool calls run in the background while the conversation continues.
• It says all audio generated by Google includes a SynthID watermark.
• The announcement does not give per-token prices. It describes pricing only as competitive.

Anthropic: a misuse report published in the same week

Anthropic grouped misuse activity it had disrupted between December 2025 and August 2026 into seven areas. Its news listing gives the report’s publication date as September 10.

One case that stands out from a parent’s perspective concerns dating-app fraud. The report says an app developer in China built a network of more than 20 dating apps and advertised human interaction while AI actually handled the conversations. Over two weeks in April 2026, more than 4,700 distinct AI personas talked to at least 25,000 people, exchanging about 2.36 million messages. It also says real people were brought in for steps intended to verify someone was human, such as video calls or social-media checks.

Another passage concerns voice directly. In one influence-operation case, the report says the operators cloned a real activist’s voice from private messages, then impersonated that person in live conversations with contacts inside Iran.

What this could mean for users

The points above describe what the announcements said. The following are our interpretations.

• Everyday users: It is becoming harder to tell whether an incoming caller is human or AI. A natural-sounding voice is becoming weaker evidence that a person is speaking.
• Parents: It becomes harder to check whether an app’s claim that children are talking to people is accurate. The case Anthropic described showed a gap between advertising and reality.
• Developers: Both companies emphasized uninterrupted conversation. The competitive focus of voice products appears to be shifting from recognition accuracy toward how natural the conversation feels.

What we could not verify

• We did not try the models ourselves. We did not run either voice model or measure its performance. Every figure above was reported by the company in its own announcement.
• Evaluation conditions: We could not verify the conditions behind each score or whether a third party reproduced the results.
• Korean-language performance: The announcement does not explain how Korean is included among the 97 languages or how well it works.
• Actual cost: OpenAI’s $0.05 per minute covers only the voice layer, so we do not know the total once the model behind it is included. Google did not disclose numerical prices.
• Report publication date and time: Anthropic’s report body had no date, so the September 10 date comes from the company’s news listing.

Sources

• OpenAI, Build more natural voice experiences with GPT-Live-1 in the API (2026-09-10) https://openai.com/index/introducing-gpt-live-1-in-the-api/
• Google DeepMind, Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking (2026-09-15) https://deepmind.google/blog/introducing-gemini-3-8-live-and-3-8-live-extended-thinking/
• Google Blog, the same announcement https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/
• Anthropic, Detecting and countering misuse of AI: September 2026 https://www.anthropic.com/threat-intelligence-report-september-2026
• Anthropic news listing (source for publication date) https://www.anthropic.com/news