Edition #11. This week Anthropic released Claude Opus 4.7. One day later came Claude Design, a design product that under the hood just runs on Opus 4.7. And on the same Thursday OpenAI launched GPT-Rosalind, a model not meant for everyone but for a very specific group: scientists. Three separate updates, but they all show the same thing about what the big labs are now focused on.
A big update from Anthropic: Opus 4.7 is out
Anthropic released Claude Opus 4.7 this week. The first reactions are strong, especially on coding, tool use and computer use.
What stands out is that Opus 4.7 not only scores better than Opus 4.6, but seems to get especially stronger on work that actually costs companies time: coding, tool use, computer use, finance and visual reasoning. Those are exactly the categories where a model needs to not only know something, but actually do something.
In practice that means: a developer who gets from bug to first working fix faster, a team that can turn loose screenshots, documents and customer questions into a usable answer faster, or an employee who has to do less repetitive on-screen work by hand. That's ultimately more relevant than a few extra points on an abstract benchmark.
What also stands out is that Mythos Preview already sits above Opus 4.7 on multiple points. Mythos is the more powerful model that Anthropic has so far only deployed with a handful of partners, mostly for security research. That suggests Anthropic is internally already a step ahead of what is publicly available.
Everyone is waiting for GPT-5.5. OpenAI seems less concerned about it than the internet
Around GPT-5.5 the same pattern as always is back: lots of rumors, lots of Reddit speculation, zero real confirmation. Dates get thrown around, codenames get tossed about, theories get spun about OpenAI deliberately waiting for Anthropic. Entertaining, but substantively it doesn't say much. It mostly says that part of the AI market now treats model releases like they're football matches.
What is interesting is that OpenAI just keeps shipping in the meantime. No 5.5 hype, but real product updates like Codex extensions and GPT-Rosalind for life sciences. Maybe that's the point: the battle is increasingly less about who tweets a new model number first, and more about who first puts something useful in place that people actually work with.
Anthropic doesn't just want Claude to think. It now also wants Claude to design
Anthropic launched Claude Design, a tool with which you can make prototypes, slides, one-pagers and other visual output by simply talking to Claude. It came literally one day after Opus 4.7, and runs under the hood on that same new model. That sounds like a logical extension, but it says something bigger: a strong model doesn't stay a standalone model for long. As soon as it's good enough, a product immediately gets built around it that teams can directly work with.
That makes Claude Design more interesting than just another model update. Not because design is suddenly solved, but because you see how quickly Anthropic moves from model to workflow. A sales deck, a first landing page or a rough product mockup soon won't have to start in PowerPoint, Figma or Canva, but just in a conversation. That's where this is heading: fewer separate tools, more directly usable work.
OpenAI isn't building GPT-Rosalind for everyone. That's exactly what makes it interesting
OpenAI came out this week with GPT-Rosalind, a model series for biology, drug discovery and translational medicine. That sounds niche, and it is. But that's exactly where the interesting part sits. Apparently OpenAI sees enough value in this kind of work to not just turn it into a general update, but something explicitly tuned to scientific workflows, genomics, protein structure, chemical reasoning and tool use in research.
That's relevant, even if you do nothing with life sciences yourself. Not because everyone now suddenly needs a biotech model, but because you see where this is going. Labs are not only building broader models, they're also building more targeted ones for work where there's a lot of value, complexity or margin for error. Today that's biology. After that you get the same pattern in legal analysis, financial controls, support work and operations.
What strikes me most is that it's increasingly less about the model itself and more about the use case. Not just: how smart is this thing? But: where do you actually get value from it in practice? That's also where my focus with AI Agents lies.
What I'm taking away
Three things stick with me. Good models get packaged into something you can directly use faster and faster. Not every important launch is for the masses, sometimes the real value sits precisely in a niche. And the difference increasingly sits in concrete work you can do faster, cleaner or with fewer people.
