Howdy wizards,

Here’s what’s brewing in AI.

Which industries build their AI implementations versus buy it?

Today, for your viewing pleasure, I’ve ranked 21 industries by build vs buy.

Source: contextwindows.ai based on 2,142 AI implementations

Pharma, Government and Insurance all build more than 60% of their AI initiatives.

Something at play here: strong regulation, strong relative importance of protecting data, good capacity of engineers that can build reliable tools.

However, at the other end of the spectrum (those that buy rather than build) we have possibly the most regulated industry ever, Healthcare, where 75% of AI projects are actually bought.

The reason is there's a #1 rockstar use case: Clinical Documentation aka AI scribes. There’s a big market of vendors that have nailed the workflow, so little incentive to build this.

An AI scribe is an ambient tool that takes the audio from patient interactions and turns it into clinical notes β€” quite simple, but saves on average 2h/day on documentation time, while also allowing doctors to stay more present with their patients.

Source: contextwindows.ai, based on 92 Clinical Documentation case studies

Same logic applies to anyone whose job runs on meetings: it's a chore that gets in the way of being present.

Which is why you've probably already used a meeting note-taker. Probably several β€” Meet’s, Zoom’s, Otter’s, Fireflies…

The question is, does your note taker actually enhance your notes instead of drowning you in transcript? Pull audio directly from your computer so you don’t annoy everyone with bots joining the call? Remind you of a nutritious, crunchy breakfast staple?

If any of those is a no, you’ve opened the right newsletter today.

This week’s sponsor is my favourite AI note taker, Granola. I should warn you that it may cause addiction. Check it out at your own risk:

IN PARTNERSHIP WITH GRANOLA

You know that feeling when you leave a meeting and immediately forget half of what you agreed to?

That's not a memory problem. When you're back-to-back all day, there's simply no time to process. No time to write the follow-up.

Granola helps you become the person who actually does what they said they'd do.

You take notes during the meeting. Just quick bullets, nothing formal. Granola transcribes in the background and turns those notes into clear summaries with actual next steps.

No more "wait, what did we decide?" moments. Just clarity. And follow-through.

Free month with the code WHATPLUGIN

NEWS NEWS NEWS ❦ NEWS NEWS NEWS

The big thing

OpenAI dropped GPT-5.5. It’s better at messy, multi-step tasks, planning, tool use and can work for longer without your supervision.

Benchmarks are strong: scores 82.7% on Terminal-Bench 2.0 (up from 75.1%), ahead of Opus 4.7 and Gemini 3.1 Pro. It’s also faster while performing at a higher level versus GPT-5.4.

The model’s pricing is $5/$30 per million tokens, which is about half of other competitive models, but far higher for the Pro tier.

OpenAI also dropped three other things in the same launch:

  • Workspace Agents: a kind of evolution of custom GPTs. They’re small agents that anyone can build in plain language, and they can handle more complex workflows than GPT ever could be cause they’re powered by OpenAI’s coding model, Codex. That also means they can work over a longer period of time, connect to all your systems and tools and work in the cloud.

  • Images 2.0: their new image model. Haven’t tested it but apparently really good β€” currently ranked #1 on the Image Arena leaderboard.

  • ChatGPT for Clinicians: a version of ChatGPT designed for clinical work. It’s specialised in documentation and medical research including real-time cited answers grounded in medical literature. They also launched HealthBench Professional, a new benchmark to evaluate clinician chat tasks.

Why it matters

The frontier of AI capability just got a meaningful upgrade, and OpenAI’s new agents are making it easy for non-technical people to turn that capability into productivity. However β€” they’ve had modest success so far with earlier variations of this concept, namely plugins and GPTs, so it remains to be seen how much users actually like them.

ChatGPT for Clinicians is a clear vertical move, and OAI is making it free for verified US clinicians to capture the market. They’ll now compete head-on with established vertical AI tools like OpenEvidence.

All the small things

China

  • DeepSeek released V4. 1M-token context, architecture upgrades, optimization improvements. DeepSeek is in talks to raise at a $20B valuation with Tencent and Alibaba β€” its first funding round.

  • Moonshot AI, another Chinese AI lab, open-sourced Kimi K2.6. It’s an agentic coding model that can run 300 parallel sub-agents for 12+ hours straight. Beats GPT-5.4, Opus 4.6, and Gemini 3.1 Pro on Humanity's Last Exam (with tools) and on SWE-bench Pro. Chinese labs are increasingly shipping frontier-class coding agents at prices that are hard to ignore.

  • The White House accused China of "industrial-scale" AI theft via distillation. The framing is that they’re copying frontier models by training on their outputs. It's the cleanest argument the administration has for their next round of export controls, and it conveniently treats every Chinese open-weight release as evidence of the crime.

Industry moves

❦

You are a delight.

Dario

See which AI use cases are paying off

Most companies pick AI use cases by brainstorming internally. 90% of those initiatives fail.

I’ve created Context Windows just so you can pick the winners.

🟦 Find high-performing use cases from 2,000+ companies at contextwindows.ai, or book a demo with me

Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.

Keep Reading