Daily Briefing
26 August 2026 · 5 stories
🔥 Top story
1OpenAI unveils its own inference chip: 1.5× more efficient than today's best NVIDIA chip
Imagine an electric engine custom-built for a single car — that is the idea behind Jalapeño, OpenAI's first AI chip unveiled this week. It was designed for a single purpose — running the GPT model — and the result speaks for itself: it produces 50% more text per watt of electricity than the best NVIDIA chip on the market, and responds 3 times faster. For the average person, this means one concrete thing: when you ask a complex question to an AI agent in a few months, the answer will arrive faster and cost less to produce. It is like waking up one morning to find your car uses half as much fuel while going faster. OpenAI is announcing a small-volume rollout by the end of the year, then wider in 2027. The figures come from OpenAI itself and have not yet been verified by an independent lab. Source: https://openai.com/index/jalapeno-hot-chips-2026
2Google launches an AI agent dedicated to the world's biggest law firms
Four of the most prestigious law firms on the planet — Cleary Gottlieb, Freshfields, Weil and Williams & Connolly — have just adopted a new tool: Gemini Enterprise for Legal, Google's first AI agent tailored for the legal profession. Concretely, it reads contracts in seconds, retrieves a legal precedent from millions of documents, and knows how to draft a procedural argument. It is like giving a junior lawyer an assistant who has already read the entire case law. The big firms bill clients by the hour — if the agent does in 10 seconds what a junior lawyer does in 2 hours, the potential savings are considerable. Google enters here in direct competition with Harvey (which already serves several firms) and with Anthropic products in the same niche. Source: https://blog.google/products/google-cloud/gemini-enterprise-legal-launch-2026
3AI agents finally get their own access badge — like an employee on their first day
When a new employee arrives at a company, they get a badge that says: "here is where you can go, here is what you can do". For AI agents, that badge did not really exist — they often circulated with an over-broad pass-all. Okta, which manages badges for 20,000 companies worldwide, has just made its Agent SSO solution generally available: from now on, each AI agent has its own identity, with its own rights, and the company can know exactly what it does and switch it off remotely. The timing is no accident: Okta reveals that only 34% of companies treat their agents with the same rigour as their human employees. Concretely, this reduces the risk that an AI agent makes a costly mistake because it was given too much freedom. Source: https://www.okta.com/newsroom/press-releases/agent-sso-ga-2026
4A new startup raises $26 million to build a Google specialised in AI agents
Keenable, founded by former executives of Yandex Search and Amazon, comes out of stealth this week with $26 million in the bank and a clear ambition: to become the reference search engine for AI agents. Their index contains 100 billion documents, and it was designed to answer thousands of questions per second — much faster than Google does for a human. Why? Because an AI agent often asks hundreds of questions in a single task, where a human asks one or two. The service is free up to 1,000 searches per hour. For a startup betting on agents, no longer paying Google on every query is an immediate saving. The competition is called Tavily, Exa or Perplexity — Keenable claims to outperform them all. To be verified in the field. Source: https://keenable.ai/blog/launch-26m-seed-2026
5OpenAI dismantles a Russian influence operation — the 9th time the company has publicly flagged such a case
Imagine a fake think tank, "International Burke Institute", publishing political analyses favourable to Moscow on X, LinkedIn, Facebook, Substack and Telegram — all this content was in fact produced by AI agents run from Russia through a network of ChatGPT accounts. OpenAI has just dismantled this network, which was also distributing anti-EU articles in German. This is the 9th time OpenAI has published such a report — proof that malicious use of consumer AI is now a regular activity for several states, not an isolated incident. Russia becomes the 4th state actor documented to divert Western AI for propaganda purposes in 2026. For the average person: next time you read a particularly well-written opinion piece on social media, ask yourself who really wrote it. Source: https://openai.com/index/disrupting-russian-ai-think-tank-2026
📡 To watch
Anthropic gives Claude a real memory — no more re-explaining everything at every conversation
When you chat with Claude on your computer, then start a task on Cowork (Anthropic's agent), it will remember the context. No more having to repeat that you hate tables or that your project is called Phoenix. What you say in chat is now available to the agent — and vice versa. It is like a human assistant who, instead of having the memory of a goldfish, remembers everything you told them last week. Available by default for all paying users.
China teases its next generation of models — Qwen 4 is being prepared behind the scenes
Alibaba has released Qwen 3.8-Flash-Next in preview, but the model itself is not the point: it is the warm-up before Qwen 4, expected as the next major step in the Chinese open-weight AI. Developers thus have a few weeks to test their integrations before the official release. The Chinese cadence remains weekly — while the West digests its announcements, Tencent, DeepSeek and Alibaba keep publishing.
GLM-5.3 — the Chinese model promising free AI for all, weights opening expected 28 August
Z.ai (the company behind GLM) is due to publish the open weights of its GLM-5.3 model around 28 August. This would be the first model of this size released without licensing restrictions, which lets anyone run it on their own servers. If you are a European company concerned with GDPR compliance, this is an option worth watching closely.
Okta invents a common language so AI agents can talk between businesses
Behind the Agent SSO announcement, Okta has published an open protocol called XAA (Cross App Access). The idea: that an agent working for your bank can properly query another service's agent, without you having to give it your passwords. It is a bit like universal access badges between companies — still at the project stage, but worth keeping an eye on.
📊 Trend
Three signals converge this week and draw the same picture: AI agents are becoming an infrastructure layer in their own right. First, OpenAI shows that we can now design a chip custom-built for its own model — like a craftsman who makes shoes to the exact size of his customers. Then, Okta acknowledges that agents need an access badge, just like an employee: they enter the company, but under control. Finally, Google equips the world's biggest law firms, proving that agents are no longer a curiosity for geeks but a serious professional tool. For a company wondering where to start: start by asking yourself what are the 5 internal processes that could be entrusted to an agent without risk, and what precise rights you would give it. It is now a topic for the general management, not the IT department.