Subscribe
Share
Share
Embed
Daily AI news and research, distilled. UpNext AI breaks down the most important developments in artificial intelligence—from major industry moves to cutting-edge papers.
Welcome to the UpNext AI podcast. It's Wednesday, October 28th, 2026, and here's what matters in AI today.
OpenAI has published a case study of Retell, a platform businesses use to build voice agents for contact centers. Retell says its monthly call volume grew from 30 million in early 2026 to more than 100 million by September. OpenAI models generate responses for 80 percent of calls on the platform. Those are two different measures: total calls handled by Retell, and the share whose responses use OpenAI models.
On a call, Retell manages the audio and conversation flow while the models use a customer’s instructions to answer and request actions in connected systems. The agent can also pass a caller to a human representative with context from the conversation. Retell reports customer resolution rates above 80 percent, a measure of requests resolved rather than calls routed to a particular model.
Model choice matters when those conversations involve company rules and multiple steps. Retell’s benchmark uses difficult, multi-turn scenarios drawn from production. In its October benchmark, GPT-6 Astra ranked first for instruction following, with a score of 89.7 percent. That is Retell’s own evaluation, measuring one aspect of voice-agent performance.
The case study also describes engineers using Codex to inspect unfamiliar code, build features and investigate bugs. Retell reports an annual revenue run rate of $1.6 million per employee. It continues to hire, but says its engineering team has not grown at the same pace as the business. The case offers a useful view of the whole operation: model responses, call-handling software, integrations and human handoffs all have a role in getting a customer’s request resolved. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast.
That demand for AI services also has a hardware story. The Information reports fresh signs that Google’s tensor processing units, or TPUs, are finding customers beyond Google Cloud. Google says Oracle will become the first major cloud provider outside Google to offer TPUs from its own data centers. The Information also reports that OpenAI plans to buy the chips, even as it remains a major Nvidia customer.
Anthropic’s IPO filing gives a sense of the commitments involved. It describes a plan to spend $114.5 billion with Google Cloud, though it does not specify how much of that amount is for TPU compute. Anthropic separately estimates it could spend another $125 billion to $150 billion on TPUs with Google and other providers. Those figures describe plans and estimates, not sales Google has already booked.
Google has only recently begun recognizing revenue from TPUs sold for use in outside data centers. Its chief financial officer said most revenue from those agreements would arrive next year.
Anthropic’s hardware commitments lead into another question about its future: who will have influence over the company if it goes public? The Verge reports that Anthropic is preparing for an IPO while maintaining its stated commitment to developing AI safely. Public markets could give it access to substantial capital, alongside shareholders seeking returns.
The Verge says Anthropic hopes its governance structure will help it resist pressure to cut corners on safety. That makes governance a business issue as well as a statement of principle: decisions about when to release more capable products could carry costs that investors notice. The company’s stated aim is clear, but an IPO would put that aim under a different kind of financial pressure. The question for investors and customers is how Anthropic’s approach holds up when those interests conflict.
For the research note, consider a team of AI agents making a decision when each member holds only part of the relevant information. Strong individual agents may still reach the wrong answer if they fail to share what matters. Researchers introduced TeamGym to test that problem using decision tasks built from legal, finance, medical and software-engineering material. It lets them identify the facts required for a correct decision and control which participants receive them.
In tests spanning eight language models and twelve source corpora, spreading those necessary facts across participants lowered team success rates by an average of 22.9 percentage points compared with giving the same facts to one model. Requiring participants to find the relevant facts in source material, rather than handing those facts to them, lowered team success rates by a further 19.9 percentage points. The researchers also report improved final decisions after training agents on team outcomes.
This is a preprint evaluating constructed team-decision tasks, rather than a report from a deployed workplace. Its useful warning for multi-agent builders is concrete: check whether agents can retrieve and communicate the facts a shared decision depends on.
...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ...
A few more developments. The Decoder reports that Google DeepMind has introduced Genie 4, a real-time world model that simulates interactive 3D environments with sound and talking characters.
The Financial Times reports that advances in AI mathematics are prompting a rethink of cybersecurity preparations that have long focused on quantum threats. Its report raises a concern for encryption planning; it does not establish that an encryption system has been broken.
And TechCrunch reports that Apple is rolling out iOS 27.2 for iPhones and iPads, adding Siri AI languages alongside security fixes.
Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers.
And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!