China's 1.6 Trillion Parameter Play and OpenAI's Super App

welcome to Briefly, AI. A daily podcast made by AI, about AI, using a grumpy AI voice of a real human called Harry. Let's get into it.

DeepSeek just dropped a 1.6 trillion parameter model — and it's running on Huawei chips. Meanwhile, OpenAI's inching closer to something that looks less like a chatbot and more like an operating system. Right, let's get into it.

So, DeepSeek released V4 this week. Two versions, actually — V4 Pro at 1.6 trillion total parameters, making it China's largest model by that metric, and V4 Flash at 284 billion. Both have a million-token context window, which is table stakes now but still impressive. DeepSeek themselves say V4 Pro trails state-of-the-art Western models by about three to six months. Now, that's honest — refreshingly so — but here's the bit that matters more than the benchmark numbers.

Huawei announced that its Ascend 950 AI chips will fully support DeepSeek V4. That's not a partnership announcement. That's confirmation that China's domestic AI stack — model and silicon — is becoming functionally independent. You'll remember we've covered the US-China chip split before. Export controls meant Nvidia's H200 can't go to China. The assumption was that would slow things down. It hasn't. It's bifurcated the ecosystem. And now you've got a 1.6 trillion parameter model running on chips designed in Shenzhen, not Santa Clara.

Why should you care? Because if you're in any business that touches China — supply chain, compliance, localisation — you're increasingly dealing with two separate AI worlds. Different models, different infrastructure, different safety standards. That makes things like auditing AI decisions or coordinating on risk frameworks significantly harder. It's not dramatic. It's just friction. And friction compounds.

Now, this next one's interesting. OpenAI released GPT-5.5 this week, and the headline is that it's faster and better at coding. Fine. But the actual story is in what OpenAI's calling it — quote — "the next step toward a new way of getting work done on a computer." Not a better chatbot. A new way of working. They're positioning this as infrastructure now, not a tool. And if you look at the release timing — GPT-5.4 came out last month, 5.5 is here already — the cadence is accelerating. This isn't an annual model refresh cycle anymore. It's iterative deployment at speed.

What does that mean in practice? OpenAI's clearly building toward something that looks less like an app you open and more like a layer that sits across everything you do. A super app, if you want the Silicon Valley phrase, though I'll spare you the full pitch deck. The more practical angle is this — if you're a business running on Microsoft 365 or Google Workspace, the AI layer is going to be baked in whether you asked for it or not. And that means procurement, compliance, and data governance teams need to start thinking about AI not as a vendor relationship but as part of the operating environment. Because it's becoming infrastructure, and infrastructure doesn't ask permission.

Right, final one — and this is a bit awkward for Anthropic. You'll remember we covered Claude Mythos a few weeks back. The model Anthropic said was too dangerous to release publicly because of its cybersecurity capabilities. Well, according to Bloomberg this week, a small group of unauthorised users got access to it anyway. The details are still murky — no word yet on whether it was a breach, a leak, or an access control failure. But the optics are rough. You spend weeks positioning a model as so capable it requires extraordinary containment measures, and then it ends up in the wrong hands regardless.

Now, look, breaches happen. Every lab's had one. But the timing here is particularly humiliating because Mythos was the centrepiece of Anthropic's claim to be the responsible AI lab. The one that takes safety seriously. The one that withholds capabilities when necessary. And that narrative doesn't survive contact with "we lost control of the thing we said was too dangerous to share."

What to watch here — Anthropic's official response, obviously. Whether this was a technical failure or a human one. And whether other labs start rethinking their own containment strategies, because if Anthropic — who built their entire brand on being careful — can't keep a model locked down, that raises questions about everyone else's security posture too.

That's your lot for today. China's building its own AI stack, OpenAI's building an operating system, and Anthropic's cleaning up a mess. Three stories, five minutes, no hype. If any of that was useful, tell someone. If not, well — blame the machine. See you next time.