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
Claude Opus 4.7 & Codex Computer Use — The Agent Race Heats Up
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 April 17th, 2026, and today we've got a big one. Two of the biggest AI labs just dropped major releases on the same day. Let's start with Anthropic. They released Claude Opus 4.7 yesterday, and it's a significant upgrade. The headline here is Software Engineering. Anthropic says users can now hand off their hardest coding work, the kind that used to need constant supervision and trust the model to handle it. It's better at following instructions, which actually means some prompts written for older models might behave differently now. Opus 4.7 takes things more literally. Vision got a major upgrade too. The model can now see images up to about three and three quarter megapixels, more than three times what previous Claude models could handle. That's useful for computer use agents reading dense screenshots or extracting data from complex diagrams. There's also a new cyber verification program for security professionals who need the model for legitimate vulnerability research. Pricing stays the same: 5 US dollars per million input tokens, 25 US dollars for outputs. But OpenAI didn't let Anthropic have the day to themselves. They announced a massive expansion of Codex that pushes it way beyond just writing code. The big feature is background computer use. Codecx can now control your Mac with its own cursor, clicking and typing in other apps while you keep working. Multiple agents can run in parallel. There's an in-app browser now too, so you can point at a web page and tell Codex what to do with it. The model can generate images using GPT, Image 1, 5, and there are over 90 new plugins connecting to tools like Jira, Circle CI, and GitLab issues. They added memory, so Codex remembers your preferences and corrections across sessions. And you can schedule automations to run future work, essentially having the agent wake itself up to continue long tasks. This is clearly OpenAI's answer to Claude's computer use capabilities, and the feature race between these two is getting intense. Alright, let's talk robotics. Physical intelligence. That two-year-old startup that's raised over a billion dollars just published research that genuinely surprised even their own researchers. Their new model, Pi Zero, 7 that's Pi 0.7, can do something called compositional generalization. In plain English, it can perform tasks it was never explicitly trained on by combining skills it learned in different contexts. In one demo, the model successfully used an air fryer to cook a sweet potato, despite having almost no relevant training data on air fryers. The researchers found just two vaguely related episodes in their entire dataset. The model somehow synthesized those fragments into functional understanding. When the researchers first tried it with basic instructions, success was 5%. After spending half an hour refining how they explained the task, it jumped to 95%. The startup is reportedly in talks for a new funding round that would value them at 11 billion US dollars. This is the kind of generalization breakthrough that could actually change how we think about deploying robots in new environments. OpenAI also released GPT Rosalind, a specialized model for life sciences research. It's built for biology, drug discovery, and translational medicine. Named after Rosalind Franklin, the model outperforms GPT-5, 4 on scientific benchmarks and comes with a life sciences research plugin connecting to over 50 scientific databases and tools. They're working with Amgen, Moderna, and the Allen Institute. The model is available through a trusted access program for qualified enterprise customers. The idea is to accelerate the early stages of drug discovery, which currently takes 10 to 15 years from target identification to approval. A few more things worth knowing about today. Quen open sourced Quen 3, 6, 35B, A3B, a mixture of experts model with 35 billion total parameters but only 3 billion active at any time. Despite that efficiency, it's competitive with models several times larger on agentic coding benchmarks. Factory, an enterprise AI coding startup, hit a$1.5 billion US dollar valuation. InsightFinder raised$15 million to help companies debug AI agent behavior, which is becoming a real need as agents get deployed more widely. Depel expanded from text translation into voice translation. Roblox's AI assistant can now plan, build, and test games more autonomously. And Canva's AI Assistant gained the ability to call multiple tools to complete design tasks. That's it for today. A lot happening in the model and agent space. See you tomorrow.