PRODUCT

How OpenAI's voice assistant got more natural

While native device assistants have long left users wanting more, OpenAI's voice models have quietly been getting closer and closer to the assistants we've always been promised. The latest version of ChatGPT Voice just took another big step in that direction.

On Wednesday, the company launched GPT-Live, a new generation of voice models that aims to make talking with AI feel even more like having a conversation with a human. At the heart of the upgrade is a new full-duplex architecture, which enables simultaneous listening and speaking, reducing latency and creating a more natural experience. 

When you are speaking to it, it can say phrases that mimic engagement, such as "mhmm," engage in quick back-and-forth banter, or even give you more time to think without interrupting. Even as GPT-Live handles more complex queries, it will continue to talk to you to maintain the flow of the conversation. 

It can do so through the second architectural change, which decouples the voice model from deeper work. As a result, when a query requires deeper reasoning, GPT-Live offloads that request to another model, GPT 5.5 at launch, and then processes it in the background so that it can keep the discussion going with you.

For instance, in a demo given to the press ahead of the release, the model performed real-time, simultaneous translation into Hindi while the presenter was speaking, then chimed in with the completed Hindi translation. This was a standout feature because, in the past, while it was processing, it made an unnatural thinking sound, causing an unnatural pause and forcing you to stop the flow of conversation.

All of these updates contributed to GPT-Live outperforming Advanced Voice Mode, its previous technology launched with GPT-4o, in both human-preferred conversations and technical benchmarks, such as scientific reasoning and agentic search, according to OpenAI. 

Starting today, OpenAI is rolling out two versions of GPT Live: GPT-Live-1, powering ChatGPT Voice for Go, Plus, and Pro Users, and GPT-Live-1 Mini, the default for free users. They will also be available in the API soon, and interested users can fill out a form to get access. 

Ultimately, one of the biggest roadblocks to using AI is its reliance on text responses, which can add friction to everyday routines. This makes GPT-Live important for broadening AI's impact, as it enables users to interact with AI more seamlessly. The predecessor was already helpful, and I even mapped my iPhone's action button to ChatGPT's Advanced Voice Mode to use it regularly throughout the day. Still, there were quirks that made it obvious you were talking to AI. But it also raises other challenges because it sounds so human-like; since something that sounds more human is more likely to be believed, and it still has the same hallucination issues as every other AI model.

TOGETHER WITH CONVEX

Code Faster, Easier, And Less Buggy With Convex

You’re probably already familiar with Convex, the backend building platform that helps your AI agents excel – but their latest updates are taking things up a notch. Convex recently launched plugins that allow you to connect Claude Code, Codex, GitHub Copilot and more directly to your Convex deployment, allowing them to…

• Read data and run functions entirely on their own

• Optimize your project’s performance via production insights

• Improve safety by limiting access to production

There’s even more updates coming in the near future, but for now, see how you can take your agents’ work to the next level with the latest from Convex right here.

GOVERNANCE

AI is now a key chess piece in US-China relations

If you needed further evidence of how powerful AI models have become or that they've risen to the level of national security concerns, then the recent actions of the world's two largest economies should be enough to convince you.

The US government restricted international access to Anthropic's Mythos and OpenAI's GPT-5.6, and now China is looking to do the same for the most advanced models from its frontier AI labs. 

When Washington realized that Claude Mythos could be used by attackers to find software vulnerabilities and exploit systems faster than anything we've seen before, it took the drastic step of issuing export controls on the software, which essentially locked it down for three weeks while the US government and major US corporations had time to harden their systems. 

According to a Reuters report, in June the Chinese Ministry of Commerce held meetings with AI labs, including Z.ai, Alibaba, and ByteDance, to discuss restricting international access to the most advanced AI models produced by Chinese labs. The government officials also raised the possibility of limiting international investors' ability to fund Chinese AI startups. 

The Chinese move could have ripple effects across the US, given the widespread use of open-weight Chinese models by US corporations and AI enthusiasts. A report from OpenRouter found that over 30% of the tokens used on Chinese models came from the US, and that number has surged as high as 46% during some weeks in 2026. 

With AI costs soaring in many organizations, IT departments and workers have turned to open-weight models since you can download them and run them locally to save costs, improve performance, and maintain privacy and security. And the Chinese open-weight models like DeepSeek, Qwen, and GLM are generally considered the most advanced. 

If China moves to restrict US access to these models, it could boost other open models made in the US, such as Nvidia Nemotron, Google Gemma, OpenAI GPT-OSS, and others. 

Meanwhile, the rhetoric continues to escalate. On Wednesday, the Chinese Ministry of Industry and Information Technology warned that Anthropic's Claude Code "contains a security back-door vulnerability that poses a serious threat." Claude Code is not officially available to use in China, but Chinese software developers have publicly stated that many use it, according to a CNBC report. And that appeared to be backed up by the fact that Alibaba banned its workers from using Anthropic tools on Monday.

Before we get too overreactive to the US-China rhetoric, let's keep in mind that none of the recent activity is likely to stop Chinese and American technologists from using each other's models anytime soon. For example, Anthropic's Mythos is now widely available to the public. The US government's restriction simply gave agencies and corporations time to prepare their cybersecurity posture before the model became more widely available. And the Chinese government is discussing restrictions on the most advanced models. The vast majority of open-weight models would likely remain available and wouldn't affect most users. In the same way, all previous-generation Anthropic and OpenAI models remained publicly available even as the government forced the companies to temporarily restrict broader access to their latest models. In other words, American-Chinese cooperation will quietly continue beneath the surface of the latest escalation in rhetoric.

TOGETHER WITH COMP AI

AI is automating everything. Except the one thing killing your enterprise deals.

It's compliance.

Your pipeline is full. Your product is ready. But the moment an enterprise prospect runs a security review, the deal stalls. SOC 2 isn't optional anymore - it's table stakes.

Comp AI's AI agents automate the whole thing: pulling evidence from 580+ integrations, generating policies for your stack, running continuous monitoring.

See How It Works

Open source, fully auditable, and built for the AI era. Startups like Dub, Corgi, and Better Auth use Comp AI to stay permanently audit-ready.

Book a Demo

RESEARCH

Study: 38% open up to AI about relationships

Emotional support and advice have long been difficult to come by, prompting a growing number of people to turn to AI for more help.

The Oxford Internet Institute surveyed 2,000 UK adults online to learn about their interactions with large language models and found that almost one-third (31%) of regular users reported using these models for emotional support, including talking through problems or making decisions. Furthermore, one quarter of regular LLM users look to these tools for meaningful conversation. 

This trust even extended into more sensitive topics, with 38% turning to LLMs for advice on personal relationships and two-thirds for information on health issues. The demographics of the respondents tended to skew the level of comfort and reliance on using chatbots for different tasks: 

• Younger users: more likely to use LLMs across all types of activity

• Women: more likely than men to use LLMs for personal and emotional support

• Men: turned to AI for help with practical tasks to a greater extent than women

Despite the differences, the overall consensus was positive: 75% of the respondents were enthusiastic about the possible benefits of AI chatbots. It is worth acknowledging that there are risks associated with using AI for sensitive issues that have traditionally been handled by humans, given the emotional intelligence and social skills AI cannot replicate. 

"Our study provides an important starting point for understanding LLM usage trends, but substantial further research is needed to explore the wider societal implications associated with LLM use for social and emotional support, including when such interactions may complement existing sources, and when they risk replacing important human relationships," said Florence Enock, Ph.D., Senior Research Fellow at the Oxford Internet Institute and lead author of the report.

Existing research from the Human-Centered Artificial Intelligence Institute at Stanford University found that AI therapy chatbots can contribute to harmful stigma and dangerous responses and are, as a result, less efficient than human therapists. That's not even factoring in the potential privacy risks involved with giving a chatbot sensitive personal information such as the state of your mental health.

Chatbots are becoming increasingly skilled at understanding natural language and mimicking the human experience with greater consistency. This makes it easy to personify or give human attributes to models that are just analyzing patterns and repeating them, which can appear to look like consciousness. Combined with the fact that AI labs are giving users free access to even their most advanced models, this creates a recipe for users to develop a dependency on AI for all sorts of tasks, including personal or sensitive ones. However, using any chatbot for assistance with mental health should be handled very carefully, both because of the privacy concerns and because of the unreliability of LLMs due to hallucinations and sycophancy. 

LINKS

• Apple reaches a deal with Broadcom to spend $30 million on US-based chips

• Meta rolls out update to turn off the camera on glasses if privacy light is tampered with

• OpenAI development arm acquires applied AI firm Northslope

• Chip startup SambaNova raised $1 billion at an $11 billion valuation 

• Mistral AI launches Robostral Navigate, model that helps robots navigate environments

• Chinese AI lab MiniMax plans to launch a 2.7-trillion-parameter model

• Claude Cowork: It is coming to mobile and web, with beta rolling out over the next several weeks, starting with the Max plan, with more plans to follow.

• Manus: Google Drive connector got upgraded to Google Workspace

• BytePlus: Dola Seedream 5.0 Pro API is now available

• GPT-5.6 Sol: Launches on Thursday 

• Grok 4.5: Cursor and SpaceXAI released its "most intelligent model"

• Capital One: Applied Researcher II (AI Foundations, LLM Core and Agentic AI)

• Basis: Revenue Operations Leader

• University of Chicago: AI Application Developer

• Teradata: Principal Engineer, Agentic AI

GAMES

A QUICK POLL BEFORE YOU GO