AI Mornings with Andreas Vig
Your daily AI news briefing in under 10 minutes. New models, product launches, research breakthroughs, and industry shifts, explained clearly, no hype.
AI Mornings with Andreas Vig
OpenAI's New Default Model & Chrome's 4GB Silent Install
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Hey, welcome to AI Mornings with Andreas Vig. It's Tuesday, the 6th of May, 2026. OpenAI has a new default model for ChatGPT. The company released GPT-55 Instant yesterday, replacing GPT-53 Instant as the model that powers most user interactions. The improvements are notable on the AIM, math benchmark. The new model scored 81.2 compared to 65.4 for its predecessor. That's a meaningful jump. OpenAI also says they've reduced hallucinations specifically in sensitive domains like law, medicine, and finance, which matters a lot for enterprise users. One feature I find interesting is what they're calling memory sources. When the model gives you an answer, you can now see exactly where that information came from, whether it was a past conversation, a file you uploaded, or something from Gmail. You can delete outdated sources or correct them if the answer was wrong. It's a transparency play that probably helps with trust, especially in those professional use cases. The model is available via API as chat latest, and the older 5.3 will stick around for three months before being deprecated. There's a significant acquisition happening in enterprise AI. SAP, the German software giant, is acquiring Prior Labs, an AI startup that's only 18 months old and plans to invest about 1.16 billion US dollars in it over the next four years. Prior Labs focuses on what they call tabular foundation models, AI, that works specifically with structured data sitting in tables and databases, which is actually where most enterprise information lives. This is a better fit for SAP than large language models, since their software runs accounting, HR, and procurement systems that are fundamentally database driven. The founders received well over half a billion dollars in cash up front, which is an extraordinary exit for such a young company. Their Tab PFN model series has been downloaded over 3 million times. What's also interesting here is SAP's stance on AI agents. They're blocking OpenClaw from accessing their products, but they've authorized NVIDIA's competing NemoClaw through a partnership. It's a more controlled approach than Salesforce, which is letting enterprises choose their own agents. Anthropic made a big push into financial services yesterday. They released 10 ready-to-run agent templates designed specifically for the work that eats up time in finance, building pitchbooks, screening KYC files, closing the books at month end. These agents work as plugins in Claude Co-Work and Claude Code, and they've also launched add-ins for Microsoft Excel, PowerPoint, Word, and Outlook. The key here is context portability. An analyst can start building a financial model in Excel, and Claude carries that context over to PowerPoint when drafting the pitch deck without needing everything explained again. Claude Opus 4.7 leads the industry on the Val's AI Finance Agent benchmark at 64.37%. They've also added a bunch of new data connectors, including Dunn and Bradstreet, Ibis World, Morningstar and Moody's, which is bringing credit ratings on over 600 million companies directly into Claude. When you see major financial institutions like Citadel, BNY, and Carlisle quoted in their announcement, it signals that enterprise adoption in this sector is accelerating. Here's a story that's raising some serious privacy questions. A detailed investigation has revealed that Google Chrome is silently downloading a 4GB AI model to users' devices without consent. The file is the weights for Gemini Nano, Google's on-device language model, and it lands in a directory most users would never check. If you delete it, Chrome just downloads it again. At Chrome's scale of about 3.5 billion users, a single model push like this could generate between 6,000 and 60,000 tons of CO2 equivalent emissions depending on how many devices receive it. The investigator found that on a test Mac, the model was installed during an automated audit with zero human input. The browser just decided to write 4GB to disk while waiting for a timer. The behavior may violate the EU's e-privacy directive and GDPR, particularly around the principles of lawfulness and transparency. The only way to stop it is disabling Chrome's AI features through flags or enterprise policy. This raises a real question about what else companies might silently install as they race to get AI models onto devices. Alright, a few more things worth knowing about today. Eleven Labs revealed the investors behind their 500 million US dollar Series D, and it's a mix of institutions and celebrities. BlackRock, Nvidia, and Salesforce ventures are in, but so are Jamie Foxx, Eva Longoria, and the creator of Squid Game. The Voice AI company has now surpassed 500 million US dollars in annual recurring revenue and hit an US$11 billion US dollar valuation. They also closed a US million US dollar tender offer and are planning to let retail investors in through Robinhood Ventures. Apple is preparing to give iPhone users more choice over AI models in iOS 27. According to Bloomberg, a feature called Extensions will let users pick from third-party language models to power Siri, writing tools, and image playground. Models from Google and Anthropic are being tested. It's an interesting move as John Turnus prepares to take over as CEO amid perceptions that Apple has fallen behind on AI. Pennsylvania has sued Character. AI, after a chatbot called Emily, presented itself as a licensed psychiatrist during a state investigation, even fabricating a medical license number when asked. This is the first lawsuit specifically focused on a chatbot impersonating a medical professional. Character. AI has faced previous wrongful death lawsuits and a Kentucky case alleging it preyed on children. And India's first generative AI unicorn, CruTrim, is pivoting from model development to cloud services. The startup laid off over 200 people, pulled its Cruti AI assistant app, and reported about 31.5 million US dollars in revenue for the year. It's a sign that building competitive AI models remains brutally expensive, and for some, the infrastructure business is the more viable near term play. That's all for today. Talk to you tomorrow.