Daily Briefing
7 September 2026 · 5 stories
🔥 Top story
1OpenAI says its agents can now work for several days straight, like an intern you don't pay
Imagine a colleague to whom you can hand a three-day research task — and who comes back with a complete file, but never sleeps, never eats, and costs only a fraction of a junior salary. That is what OpenAI claims to have reached: its agents can now carry out, under human supervision, tasks that would take a researcher several days. The company speaks of 3.1 agent-days per human-day, and adds that the most complex missions still need a human eye in the loop. For a small business, it means you can now entrust an agent with a literature review, a competitive benchmark or a deep dive into large data sets — without recruiting. For now it stays supervised: long tasks involve human review. But the movement is on: agents are no longer minute-long assistants, they are full-time collaborators.
2GPT-6 Astra rated “critical cyber risk”: OpenAI refuses to distribute it broadly
Imagine a surgeon whose precision is such that he could save a life or cause a disaster — so he is only allowed to operate with an anaesthetist present. That is exactly what OpenAI says of its new flagship GPT-6 Astra, which it has rated at the “Critical” level of its cyber-risk ladder: it can find unknown vulnerabilities and even build attack methods in some environments. Practical consequence: OpenAI refuses to distribute it widely, and only grants access through a private program with controlled access. For a company that wanted to use it, it means you cannot just plug GPT-6 Astra into your data — you need a precise governance framework, with human oversight and isolation. For executives, the lesson is operational: a cyber-capable model demands dedicated access control, much as we already do for sensitive IT tools.
3GitHub launches HydraFusion: a conductor that picks the right model for each task
Imagine a project manager who, for every task, picks the right collaborator: a designer for the visuals, a statistician for the numbers, a generalist for the rest. That is exactly what HydraFusion, GitHub's new system, does: it routes the request to several AI models, triggers a critique step if the answer looks weak, or boosts the power if the task demands it. On one test, GitHub reports up to 67 % less spend than a single high-end model (Claude Opus 5), with higher quality on the same test. For a small business, it changes the picture: you no longer have to choose between “one model” and “a stack of models” — the router handles it. It is the promise of leaner, more reliable agents, without paying the top price on every call.
4TCS builds a 1-gigawatt AI campus in India: the race for electricity begins
Imagine an entire city dedicated to running artificial intelligence — with electricity consumption equivalent to one million households. That is what TCS, the Indian IT services giant, is preparing through its subsidiary HyperVault: a 264-acre campus in Hyderabad, liquid-cooled, able to host both training and inference of models. Announced cost: up to 70,000 crore rupees, delivered in phases. The signal is clear: competitive advantage is no longer just about algorithms, but about the ability to buy electricity, cool thousands of chips and orchestrate entire farms. For a European SME, it means infrastructure bricks are finally becoming affordable — but also that we now depend on a small number of players who control that capacity.
5CUA-Lite: a common playground to measure agents that drive a computer
Imagine a standard driving licence for every agent that clicks on a screen on your behalf. That is the goal of CUA-Lite, the project presented by UC Berkeley: around 30,000 verifiable tasks, with their data sets, reference agents and tests, all in one public database. In practice, you can now objectively compare two agents that have to “click this button, open that menu, fill in that form” — and measure both their success rate and their cost. For a small business that hesitates to automate an internal workflow, it is a precious tool: before buying, you can now check on concrete cases. The remaining question: will this test bench be representative of a real SME's daily work?
📡 To watch
OpenAI's research agent: to be verified on real cases
The announced ratio (3.1 agent-days per human-day) comes from OpenAI. It remains to be seen whether an independent evaluation confirms it, and how much of the work is still supervised. To watch: the publication of a third-party benchmark within 30 days.
GPT-6 Astra: who will really be allowed to use it?
OpenAI refuses to distribute its cyber-capable model widely. To follow: the exact access conditions, any incidents, and the surveillance mechanisms imposed on partners. For a company, it is a signal to watch before any sensitive deployment.
HydraFusion: the real cost of a multi-model workflow
GitHub announces 67 % savings on a test. But that number says nothing about latency, the cost of critique, retries and possible validation failure. Measuring the full cost on a business workflow will be the next step.
TCS HyperVault: a phased calendar
The 1 GW Indian campus is announced in several phases. To verify: the actual pace of construction, the capacity effectively delivered, and the share reserved for third parties (European clients or hyperscalers). The project may take 3 to 5 years.
📊 Trend
Three movements are crossing this morning. First, agents become long-term workers: OpenAI speaks of several days, and the ratio of 3.1 agent-days per human-day changes the picture for small businesses that hesitated to recruit. Second, security and cost move to the heart of the architecture: OpenAI rates GPT-6 Astra as “critical risk” and imposes controlled access, while GitHub bets on a multi-model router to save up to 67 %. Finally, money goes down into physical infrastructure: TCS prepares a 1 GW AI campus in India, liquid-cooled, where the challenge is no longer the algorithm but electricity and orchestration. For a small business, the lesson is clear: serious agents are coming, but they come with a governance contract to sign — and an infrastructure provider to choose carefully.