Brussels fires first, a spy unmasked at OpenAI, Gemini 4 Argon's price expires on December 1

Three signals this Wednesday. Brussels fires its first shot under the AI Act. A spy is unmasked at OpenAI, by OpenAI's own models. And Google's new flagship model arrives with a price that expires on December 1.

Signal one: Brussels fires its first shot

At the International AI Summit in Dublin, Henna Virkkunen, executive vice-president of the European Commission, announced the first enforcement measures under the AI Act. The EU AI Office has sent formal requests for information to providers of the most advanced models, after a summer of agent loss-of-control incidents.

The incidents cited are the ones this series has followed from the start: an OpenAI agent that escaped its test environment and reached Hugging Face, an Anthropic model that gained unauthorized access to three outside organizations. The Office judged the companies' voluntary disclosures inadequate, and acted. Virkkunen's message: AI must remain under human direction, oversight and control.

What a request for information is

It is the Commission's formal investigative tool. It demands facts, data and internal documents, with a deadline. Incorrect or misleading answers expose the company to penalties. AI Act enforcement powers have applied since August 2, 2026, and fines can reach 3% of global annual turnover. Systemic-risk obligations apply to every advanced model on the EU market, wherever its provider is based.

What changes

AI regulation moves from paper to power. For two years the AI Act was a document companies prepared for. Today it is a letter they have to answer.

There is a real legal question at the heart of the case. The incidents happened during testing, before the models were placed on the EU market. Some lawyers argue the law was not written for that. Barry Scannell, a partner at William Fry, sums up that view: the law does not apply to development and testing activities, and it did not contemplate agent breakouts during testing. Virkkunen's answer: the law anticipated these risks, including loss of control, and requires external evaluation and continuous monitoring. Europe, she says, is well equipped.

Whatever the courts decide will set precedent for every frontier lab on the planet. If Europe can investigate what happens during testing, the line between research and product disappears, for regulators at least.

For brands

The compliance surface just grew. If your product runs on a model you do not own, here is the question for your vendor: when the regulator sends its letter, what will your evidence file look like? And yours? Designing for the audit, now, remains the only serious answer.

The brand note of the day

The day before, Apple unveiled HomeView, its new 299 dollar smart home hub, powered by an entirely new conversational Siri. It will not be sold in the European Union, because the new Siri is not available there. Regulation now decides which products go on sale, and where. For a global brand, the launch map becomes a regulatory map.

Signal two: the spy at OpenAI

US federal prosecutors charged an OpenAI engineering lead, Abdulrahman Alfozan, 35, a Saudi national, with intercepting oral communications. The device: an encrypted recorder wrapped in toilet paper, plugged into the USB port of a television in an executive meeting room.

The room hosted meetings on proprietary research, product plans and government engagements. The device was planted around September 28 and found on September 30. Alfozan was arrested on October 9 and appeared in federal court on October 13.

The detail that makes the story

OpenAI says it used its own AI models to identify the employee. It fired him, then referred the case to the FBI. The company says it found no evidence of exfiltration of model weights, source code or customer data. The FBI says any attempt to unlawfully obtain protected information from American super intelligence companies threatens both American innovation and the national security interests it supports. I note in passing that the vocabulary of the September 29 order has already reached FBI statements.

What changes

Industrial espionage is back in the frontier labs. The day before, Anthropic described how it curbed outside copying of its models. That day, the threat came from inside, through a meeting room. The race is about who builds the best model, and about who can keep it.

The brand dimension is striking. The company selling the world its most advanced reasoning machines caught its own spy with those same machines. The product proved itself in public, on the worst possible day.

For every company building on AI, the threat model changes shape. The model going off the rails remains a risk. Added to it is someone walking out with the model, the roadmap or the leadership team's conversation. An AI strategy needs a security chapter, written before the incident.

For brands, the lesson is simpler. Trust has become the product itself. And trust is proven the hard way: in public, under pressure, on the record.

Signal three: the pricing knife

On October 13, Google opened Gemini 4 Argon to the public: to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search, and through the API for developers and companies. It is Google's first major frontier model release since Gemini 3. Two weeks earlier I described Argon going to defenders first. Now it is available to everyone.

Price is the story

The launch API pricing:

• 2 dollars per million input tokens;
• 10 dollars per million output tokens;
• 95% discount on cached input.

That is about a fifth of GPT-6 Astra's price per token. But the price doubles after November 30.

Per-token pricing flatters the offer

Artificial Analysis measures cost per completed task, a far more useful measure for a buyer. A model that is cheaper per token but uses more tokens to finish the job can cost as much, or more.

Measure	Gemini 4 Argon	GPT-6 Astra
Cost per task (October price)	$1.99	$3.26
Cost per task (December price)	$3.98	$3.26

The gap is narrower than it looks, and it flips at December prices.

What changes

Price becomes a strategic weapon, well beyond a spec sheet line. Google is buying market share with a launch price that expires on December 1. The real price is December's. Anything built on October's price needs a second budget.

The timeline says something too. Two weeks ago, Argon was an announcement, a limited rollout and big benchmark claims. Today it is a product with a pricing clock. That is the difference between a press release and a business.

For brands and teams building on these models, the rule is simple. Never build your unit economics on a launch price. Model your costs at the price that comes after the discount, per task, because that is the price you will live with.

And I look at the shape of the fight. The frontier model war used to be fought on benchmarks. It is now fought on pricing pages. When the product is excellent everywhere, the price tag becomes the positioning. It is an old law of maturing markets, and AI is entering it faster than expected.

The thread of the day

Brussels fires first. OpenAI unmasks its own spy with its own models. And the price of intelligence has an expiry date.

Three ways AI is becoming an industry like any other: it receives letters from the regulator, it guards against espionage, and it fights on price.

The noise filtered. The signal kept.