Frontier Models

Anthropic's Claude Sonnet 4.5 takes an agentic lead forward

Anthropic released Claude Sonnet 4.5 this week; its headline feature is that the model can work autonomously on complex tasks for over 30 hours without losing focus. To demonstrate this, Anthropic had the model code an entire Slack-like chat application—11,000 lines of code—and it kept running until the job was done. This represents a massive leap from earlier this year, when Anthropic's Opus 4 made headlines for maintaining coherence for seven hours. In the past, AI models typically lost the thread as errors accumulated and their context windows filled up.

Additionally, Anthropic reports significant improvement in Claude's computer use functionality. From The Verge:

"Dianne Penn, a head of product management at Anthropic, told The Verge in an interview that the model’s improvements in its computer use capabilities surprised even her. Claude Sonnet 4.5 is more than three times as skilled at navigating a browser and using a computer compared to Anthropic’s tech from last October."

Beyond raw performance, Anthropic also claims the model shows reduced sycophancy (the tendency to praise even bad user ideas) and "the tendency to encourage delusional thinking" compared to prior models.

Anthropic releases Claude Sonnet 4.5 in latest bid for AI agents and coding supremacy
⚡ Quick Read (3 minutes)

Agentic AI

Microsoft embraces "vibe working" in Office apps

Microsoft introduced Agent Mode for Excel and Word, alongside a new Office Agent in Copilot chat, bringing what they call "vibe working" to knowledge workers. Similar to vibe coding, where novices create apps through simple prompts, Agent Mode lets users generate complex spreadsheets and documents by describing what they need in plain language. The Agent Mode feature, powered by OpenAI's latest reasoning models, breaks down complex tasks into steps you can follow and audit in real-time, while the Office Agent uses Anthropic models to create PowerPoint presentations and Word documents from chat prompts.

Microsoft has been cautious about adding AI to Excel, given the high stakes of spreadsheet accuracy. Their own SpreadsheetBench results show why: Agent Mode achieved 57.2 percent accuracy compared to 71.3 percent for humans. As with vibe coding, the gap between AI capability and human expertise means these tools work best for lower-stakes work where experienced users can carefully audit the output.

The features are currently available through Microsoft's Frontier program for Copilot customers, with Agent Mode launching first in web versions of Excel and Word before coming to desktop apps.

Microsoft's move signals that major productivity platforms are shifting from AI as an assistant to AI as a co-creator of work artifacts. The simultaneous use of both OpenAI and Anthropic models also shows how companies are beginning to mix and match AI capabilities rather than committing to a single provider.

Microsoft launches ‘vibe working’ in Excel and Word
⚡ Quick Read (5 minutes)

Agentic AI

AI agents can now complete purchases on your behalf

Stripe and OpenAI just launched the Agentic Commerce Protocol (ACP), an open standard that lets AI assistants like ChatGPT handle purchases directly within conversations. Starting now, ChatGPT users can buy from U.S. Etsy sellers and soon from Shopify merchants without leaving the chat interface.

The protocol addresses three critical challenges that have been hindering AI-powered commerce: maintaining trust and security when AI initiates transactions on behalf of buyers, avoiding fragmentation by establishing a single standard instead of requiring businesses to build custom integrations for each AI platform, and supporting complex purchase flows beyond simple one-time purchases.

Significantly, ACP is open source and designed to work with any commerce backend or payment provider. Businesses maintain control as the merchant of record while enabling their products to be discoverable and purchasable through any AI agent that adopts the protocol. This represents commerce shifting to "buying where you discover"—and AI assistants are increasingly where that discovery happens.

Developing an open standard for agentic commerce
⚡ Quick Read (4 minutes)

AI Content Creation

OpenAI launches Sora 2 with synchronized audio and a social twist

OpenAI released Sora 2, its second-generation video AI that can now generate synchronized dialogue and sound effects. The model shows substantial improvements in physics modeling (missed basketballs now bounce off backboards instead of teleporting into hoops) and can handle complex movements like gymnastics routines.

The notable twist is the accompanying social app (currently available only on iOS), featuring "Cameos," which allows users to record themselves once and then insert their likeness and voice into any AI-generated scene. While the company has implemented safety controls—users decide who can use their likeness and can revoke access—the implications for deepfakes are significant. The app is launching in the US and Canada on an invite-only basis, offering initial free access with usage limits.

OpenAI’s Sora 2 lets users insert themselves into AI videos with sound
⚡ Quick Read (5 minutes)

AI Product Development

Lovable simplifies full-stack AI app development with an integrated backend

No-code development platform Lovable launched two major features this week that eliminate traditional friction points in building AI-powered applications:

• Lovable Cloud provides an integrated backend that automatically provisions databases, authentication, and file storage through simple prompts—no external dashboards or configuration required.

• Lovable AI adds AI functionality to apps without requiring API keys or managing multiple billing relationships, automatically connecting to models from Google Gemini and OpenAI.

The launch demonstrates how infrastructure is being abstracted away from builders, letting them focus on product features rather than technical setup. The demo video features a 10-year-old boy creating a fully functional app using only prompts.

Both features follow usage-based pricing with generous free tiers.

Introducing Lovable Cloud and AI
⚡ Quick Read (5 minutes)

That’s it for this week.

Thanks for reading, and see you next Wednesday with more curated AI/UX news and insights. 👋

All the Best, Heidi