A lot happened this week in AI - and not all of it was hype. We saw OpenAI drop what might be their most powerful model yet, Meta quietly release a frontier model that's turning heads, and Shopify prove that smaller, smarter models can beat the big guys on specialized tasks. Meanwhile, IBM's Watson Health failure resurfaced as a cautionary tale, and the entry-level job market is showing real cracks under AI pressure. Here's what actually matters.

Summary: OpenAI released GPT-6 Astra, its most capable model to date, designed to operate computers and handle long-running tasks. It scored 72.6% on real-computer tasks and hit 99.9% on ARC-AGI-3 - up from GPT-5.6 Sol's 7.8%. OpenAI's president Greg Brockman said Astra may qualify as true AGI, though he's leaving that for users to decide.

Why it matters: This isn't just a benchmark flex. Astra can run software, manage files, write and debug code, and stay on task for extended periods - things that used to require a human in the loop. Access is rolling out slowly due to cybersecurity concerns, but when it's widely available, it could change how knowledge work gets done.

Key takeaways: GPT-6 Astra outperforms GPT-5.6 Sol on computer tasks and takes 40 minutes avg vs. 75 minutes. Priced at $10/$50 per million tokens - costlier, but more efficient per completed task. OpenAI classified it as 'Critical' cybersecurity risk and is rolling out access carefully.

Source

Summary: Meta released Muse Spark 1.3, a model that places it in the same conversation as Anthropic and OpenAI. It scores 62 on the Intelligence Index - behind only Claude Fable 5.1 and Opus 5 - and is significantly cheaper to run than most competitors. It's available for free on OpenCode with low API costs.

Why it matters: Meta entering the frontier model race with a cost-efficient option puts pressure on every other lab. If you don't need the absolute top-tier model, Muse Spark 1.3 becomes hard to ignore - especially for businesses watching their AI spend.

Key takeaways: Needs roughly 20% fewer tool calls and 25% fewer tokens than its predecessor. CEO Zuckerberg hinted at a larger upcoming model codenamed 'Watermelon'. Plans to release Spark's model weights publicly.

Source

Summary: Anthropic launched Claude Fable 5.1 with a 25% cost reduction for standard use and up to 45% for automated pipelines. It now ranks first on the Intelligence Index and more than doubled its previous score on scientific research benchmarks. A parallel release, Mythos 5.1, exists with fewer guardrails but is restricted to vetted researchers.

Why it matters: For teams already using Claude, this is a meaningful upgrade - lower costs, better performance, and fewer annoying refusals on legitimate questions. The 75% cut in cache read costs alone makes a big difference for anyone running high-volume AI tasks.

Key takeaways: False refusals dropped significantly on cybersecurity and medical questions. New Enterprise Frontier Safeguards let sensitive data stay on your own cloud. Scored 52.6% on Terminal-Bench-Science vs. 24.7% for the previous Fable 5.

Source

Summary: IBM's Watson Health was supposed to revolutionize cancer treatment. Instead, it recommended questionable treatments, struggled with real patient data, and required constant fact-checking by hospitals. In 2022, IBM sold it to a private equity firm for a fraction of what it spent - a cautionary tale about overhyping AI before it's ready.

Why it's interesting: This is a good reminder that flashy AI demos don't always survive contact with reality. Watson crushed it on Jeopardy! but fell apart when patients had multiple conditions and messy medical records. The gap between marketing and actual capability is still very real.

Key takeaways: Watson's own lead scientist warned that pattern-matching AI wasn't suitable for oncology. Hospitals found they still had to fact-check every recommendation - defeating the purpose.

Source

Summary: Meta's Project OT aimed to make the company AI-native and cut its workforce by 60%. After the first wave of layoffs, technical incidents shot up 40%, repair times rose 70%, and AI-generated internal work was only 36% useful. Employee morale tanked, petitions circulated, and Zuckerberg cancelled the second wave.

Why it's interesting: Even Meta - one of the biggest AI investors on the planet - couldn't just swap out humans for AI and call it done. This is a real-world stress test of what happens when you move too fast.

Key takeaways: AI-driven internal work increased 220%, but only a third of it was actually useful. Thousands still lost their jobs for a technology that wasn't ready to replace them.

Source

Summary: Surgeons at the National Hospital for Neurology and Neurosurgery in London used a real-time AI tool to color-code anatomy during an 11mm brain tumor removal. The AI drew on patterns from hundreds of previous surgeries to help the team avoid cutting anything that would have caused blindness. The patient kept his sight.

Why it's interesting: This is one of the clearest examples of AI actually saving a life in real time - not a chatbot answering questions, but an AI system helping a surgeon make better decisions mid-operation.

Key takeaways: The AI provided live video analysis, not just pre-op planning. Marks a shift from AI as a research tool to AI as an active part of the operating room.

Source

Claude Fable 5.1 - Anthropic's latest model - faster, cheaper, and better at coding and research than its predecessor. Now with background Mac support so it can work while you do other things. anthropic.com

OpenClaw 2.0 - An open-source personal AI agent that now supports multiplayer sessions, a rebuilt browser app, and memory features. Works with ChatGPT or Claude subscriptions. Available on iOS and Android. openclaw.org

Gemini Notebook (NotebookLM) - Upload your own documents and ask questions - Gemini answers with inline citations so you always know where the information came from. Great for research, meeting notes, and project briefs. gemini.com/notebook

Runway Solaris - A wild early-access tool that generates website interfaces frame by frame like a video - no code underneath. Still experimental, but it's a glimpse at what web design might look like in a few years. runwayml.com

Fambot - An AI assistant that pulls together your emails, WhatsApp messages, and calendar into one daily family checklist. Built for parents who are drowning in group chats and school notifications. fambot.com

Context: Inspired by Shopify's self-improving AI pipeline - the idea that real failures are the best training data. You can apply the same principle to your own workflows by using AI to critique and improve its own outputs.

Prompt: "Here is a task I gave an AI and the output it produced: [paste task + output]. Review this output critically. What did it get wrong, miss, or handle poorly? Then produce an improved version that fixes those issues. Explain what you changed and why."

Example use case: Use this when an AI draft - whether it's a report, email, or analysis - feels off but you can't quite put your finger on why. Feed the original prompt and output back into the AI and let it self-critique. You'll often get a noticeably better second version without having to rewrite your prompt from scratch.

Summary: UK top employers cut new graduate hiring by nearly 25% since 2022. In the US, employment for 22-25 year olds in AI-exposed jobs is roughly 19% below where it would otherwise be. Companies are letting AI handle the routine work that used to go to junior hires.

Why it's important: Every senior expert was once a junior who learned through repetitive, entry-level work. If those jobs disappear, companies are cutting off the pipeline that produces experienced talent - the same talent they'll pay top dollar for in five years.

Summary: OpenAI published data showing its researchers now use 3.1 agent workdays for every human workday, calling it the 'automated research intern' milestone. The same day, Chief Scientist Jakub Pachocki warned that recursive self-improvement is accelerating faster than the lab's alignment and monitoring systems can keep up. Humans still step in on more than half of successful 4-8 hour agent tasks.

Why it's important: It's a direct admission from inside the company building the fastest AI that using AI to build better AI may be outrunning the safety work meant to keep it in check. For any business leaning harder on AI agents, it's a signal that oversight needs to scale alongside adoption, not trail behind it.

Source

Summary: At least 37 Americans were arrested in 2025 protesting new data centers, including a Kansas physics teacher removed from a public meeting for applauding an anti-data-center speaker. Local opposition has contributed to blocking or delaying $130 billion worth of projects in Q1 alone.

Why it's important: The AI boom runs on data centers - and data centers need land, water, and power that local communities are increasingly unwilling to give up. This isn't a fringe issue anymore. It's a real bottleneck on AI infrastructure expansion.

That's it for this week's Weekly Round-up. Forward this to one person who wants to stay ahead of AI. See you next week. - Nicolas