研究ニュースに戻る
Editorial cover for live voice, agentic AI, and source-grounded notebook product news
政策 / 倫理Industry signal20262026年7月20日· 10 min

News: ChatGPT Voice, Claude Sonnet 5 and Gemini Notebook expand AI interaction

500語要約

Learners and an educator supervise a voice conversation, an agentic workflow, and a source-grounded digital notebook.

This product-news roundup tracks three changes to the way people interact with AI. OpenAI's GPT-Live moves ChatGPT Voice toward continuous conversation. Anthropic's Claude Sonnet 5 makes tool-using, multi-step agents more widely available. Google has renamed NotebookLM as Gemini Notebook and is connecting the source-grounded research product more deeply to the Gemini ecosystem. For AIEDHK, the shared signal is that AI is no longer confined to a turn-by-turn chat box: it can listen continuously, act across tools, and maintain a working collection of sources.

OpenAI launched GPT-Live on July 8 as the new voice-model family powering ChatGPT Voice. Its full-duplex design can listen and speak at the same time, decide whether to pause or respond many times per second, and handle interruptions more naturally than earlier turn-based voice systems. When a request needs web search, deeper reasoning, or agentic work, GPT-Live can delegate it to a frontier model in the background while maintaining the conversation. GPT-Live-1 and GPT-Live-1 mini are rolling out globally in ChatGPT, with API access planned later.

That interaction model could support oral language practice, accessible hands-free learning, think-aloud explanation, and formative questioning. But conversational fluency can also make an answer feel more trustworthy than it is. Educators should therefore design voice activities that require learners to restate evidence, compare an answer with course materials, and demonstrate understanding after the conversation. OpenAI reports preference and capability evaluations, not independent evidence that GPT-Live improves learning, language development, or accessibility outcomes.

Anthropic's June 30 release positions Claude Sonnet 5 as its most agentic Sonnet model yet. Anthropic says it can plan, use browsers and terminals, write and test code, and sustain complex work at a lower price than its larger Opus models. Sonnet 5 is the default for Claude Free and Pro accounts, is available across other plans, and is offered through Claude Code and the Claude API. Users can vary the effort level to balance cost, speed, and performance.

For education, broader access to capable agents changes the assessment problem. A student can delegate research, coding, file changes, and verification steps that once left clearer evidence of personal effort. Agentic tools can still be educational when students must define the plan, constrain permissions, inspect each artifact, explain failures, and defend the final choices. Vendor safety tests and self-checking behavior are useful product signals, but they do not remove the need for sandboxing, data controls, provenance, or teacher review.

Google's July 16 announcement brings an explicitly research-and-learning purpose into the roundup. NotebookLM is now Gemini Notebook: the standalone, source-grounded product remains focused on helping people learn and research, while gaining tighter connections to the Gemini app and, in future, Google Search. Google says more than 30 million people and 600,000 organizations use the product, including students who turn notes into audio and video summaries. A secure cloud computer is also rolling out so Gemini Notebook can write and execute code for analysis grounded in notebook sources.

For Hong Kong schools and universities, these products invite three different pilots: voice for explanation and language interaction, agents for supervised project work, and notebooks for source-grounded inquiry. Each pilot needs its own evidence. Voice should be assessed for reasoning and inclusion, not naturalness alone. Agents should expose plans, actions, and human approvals. Notebooks should preserve citation checking and distinguish uploaded evidence from model-generated interpretation. The most important news is not that interfaces feel smoother or agents finish more tasks. It is that education must make learner agency, source quality, and accountable oversight visible as AI becomes more ambient and more capable.