OpenAI launches its most powerful model and reveals bigger ambitions. Four Chinese labs release frontier models in 12 days. Google doubles down on its Pentagon deal. And the data centre buildout hits an unexpected wall.
Welcome back to This Week in AI. Episode 3 covers a week that felt like several months compressed into a few days. A landmark model launch. A geopolitical AI sprint. A governance battle inside Big Tech. And the quiet emergence of something genuinely new — AI agents that can set up an entire business on the internet, autonomously, without a human in the loop. Let's get into it.
OpenAI launched GPT-5.5 on April 24 — its most capable model yet and, according to the company, a meaningful step toward a new kind of computing. The model excels at agentic coding, knowledge work, computer use, and early scientific research — tasks where progress depends on reasoning across context and taking sustained action over time.
But the bigger story is what GPT-5.5 signals strategically. OpenAI co-founder Greg Brockman described it as a step toward a "super app" — a unified platform combining ChatGPT, Codex, and an AI browser into one service that continuously understands user context and executes tasks directly. Think of it as an AI that doesn't just answer questions — it operates software, moves across tools, and completes multi-part tasks without needing to be hand-held through each step.
GPT-5.5 also passed the same cybersecurity benchmark that Anthropic's Mythos cleared — completing a 32-step end-to-end cyber-attack range. The UK AI Security Institute now estimates frontier cyber-offence capability is doubling every four months. A sobering metric for anyone thinking these capabilities are still in the research phase.
GPT-5.5 is available now on ChatGPT Plus and above. If you use ChatGPT for work, the agentic capabilities — the ability to plan, use tools, and complete multi-step tasks — are meaningfully better than before. Test it on your most time-consuming workflows this week.
In a 12-day window in late April, four Chinese AI labs released frontier-level open-weight coding models: Z.ai's GLM-5.1, MiniMax M2.7, Moonshot's Kimi K2.6, and DeepSeek V4. All four hit roughly the same capability ceiling on agentic engineering benchmarks — matching Western frontier models — while costing less than a third of Claude Opus per inference call.
These are not research demos. They are production-ready, open-weight models with self-confident demos and real market reactions — Zhipu's stock closed up 15.92% on the day GLM-5.1 launched. China has been blocked from NVIDIA's most advanced chips by US export controls. The fact that these labs are producing competitive frontier models anyway suggests the capability gap is narrowing faster than Western observers expected. Meanwhile, China formally blocked Meta's $2B acquisition of Manus — the Chinese agent startup — marking the first state-level prohibition of an inbound AI acquisition.
For businesses building on AI infrastructure, Chinese open-weight models offer frontier performance at dramatically lower cost. DeepSeek in particular offers roughly 27× cost savings versus OpenAI for comparable reasoning tasks — a figure worth knowing if you are evaluating AI API providers for volume use cases.
Following last week's news that Google signed a classified Pentagon AI deal allowing Gemini models to be used inside US military networks for "any lawful purpose," Google's leadership has responded to nearly 600 employee signatories with a firm rebuttal — telling staff in a memo that the company "proudly" works with the US military.
This is a significant shift from the Project Maven moment in 2018, when employee protests and thousands of signatures led Google to not renew its defence contract. Current and former employees tell Fortune that the leverage workers once had has eroded dramatically — cost-cutting, AI-driven efficiency, and a weakened tech labour market have made collective action far less effective than it was five years ago.
Google has followed both OpenAI and Elon Musk's xAI in agreeing to allow its models inside classified US military networks. AI neutrality — the idea that AI companies would not take sides in defence applications — is effectively over for the major labs. The question now is not whether frontier AI will be used in military contexts, but how, under what constraints, and with what oversight.
For businesses with ethical procurement policies, understanding which AI providers have defence contracts is now a relevant vendor due-diligence question. For employees in tech, this shift in power dynamics is worth understanding — the era of influential employee activism at major AI companies appears to be closing.
Cloudflare and Stripe have jointly introduced a protocol that allows AI agents to create accounts, purchase domain names, and deploy web applications — entirely without human intervention. The system standardises identity, authorisation, and payment processes so that an AI agent can move from an idea to a live product on the internet without a single human action required.
This is not a demo or a research paper. It is live, in open beta, and available now. Separately, Adobe is testing an agentic AI inside Firefly that can execute complex, multi-step creative tasks across Photoshop, Illustrator, and Premiere — and Mistral has launched an orchestration engine designed to move AI systems from experimentation into full production business processes.
The barrier between "AI assistant" and "AI that does things in the world" has effectively collapsed. For entrepreneurs and small businesses, this means AI can now handle an increasing share of setup, infrastructure, and deployment work autonomously. For anyone thinking about building an online business — the cost and complexity of the initial setup just dropped dramatically.
The $600 billion AI infrastructure buildout — covered in Episode 2 — is running into an unexpected obstacle: local resistance. At least 11 US states have proposed restrictive data centre legislation, citing environmental concerns, water usage, energy demand, and community impact. A federal moratorium bill from Senators Sanders and Ocasio-Cortez threatens to halt new data centre construction entirely until environmental and worker protections are formalised.
Data centre NIMBYism — "Not In My Back Yard" — is now a first-order bottleneck to AI scaling. The labs anticipated resistance; they did not anticipate it arriving this fast and at this scale. One proposed solution gaining traction: data centres in space. RAAIS 2026 speaker Starcloud is already building this infrastructure — satellite-based compute that sidesteps terrestrial land and energy constraints entirely.
This infrastructure constraint will slow the rate at which AI compute becomes cheaper and more accessible — at least in the short term. For businesses planning AI-heavy workflows, locking in cloud compute contracts now at current rates may be strategically smart before supply constraints push prices up.
Three episodes in and a pattern is becoming clear. Every week brings a new frontier model, a new geopolitical development, and a new demonstration of AI doing something that previously required a human. The velocity is not slowing — if anything it is accelerating. But this week's most important story might not be any single headline. It might be the Cloudflare-Stripe announcement: the moment when AI stopped being a tool you use and became an agent that acts. That shift — from assistant to actor — is the one that will matter most to businesses over the next 12 months. The question is not whether you are ready for smarter AI. It is whether you are ready for AI that simply gets things done.