AI Recap: The week Brussels enforces the AI Act and ChatGPT starts clicking in Chrome

Gemini 4 Argon is available to everyone, but the launch price only lasts until November 30

Google made its new flagship model Gemini 4 Argon broadly available on October 13. Before that, only selected testers had access. Now you can get Argon through the paid Gemini API, Google AI Studio, the agent development environment Antigravity, and eligible Gemini Enterprise subscriptions. The price: $2 per million input tokens ($4 from December 1, 2026), cached input at $0.10 with a 95 percent discount, and output including thinking tokens at $10 ($20 from December 1). Argon is not available in the free tier.

For your business, that means you have about six weeks to test Argon cheaply on real tasks. But plan for double the price from day one, or your calculation will tip over in December. Watch out for thinking tokens: they are billed as output, so long chains of reasoning cost real money. The caching discount pays off if you send the same long instructions or manuals with every request.

ChatGPT now fills out forms in your Chrome, if you pay for it

OpenAI upgraded its Chrome extension on October 15, 2026: ChatGPT can access open tabs, summarize content, and, in Work mode, use computer use to fill out forms or compare products across multiple tabs. With @chrome you can address individual tabs directly. In Work mode, the tool controls the tabs, sends messages, fills out forms, downloads files, or groups tabs. Pay attention to the plan: The extension itself is available on all ChatGPT plans. According to OpenAI's July 9 release notes, however, Work mode is only rolling out to paid plans, with Free and Go left out. On the free plan, ChatGPT can read your tabs but won't click on its own.

For your operations, this is the real crux of the week. The agent works in your people's ordinary browser, with every login that happens to be open: CRM, webmail, supplier portals. That saves time on mindless portal work, but it moves the risk straight to the workstation. Define which sites the agent may act on, which actions need approval, and who is liable if it sends something wrong. How to set this up properly is covered in our security guide for AI agents.

OpenAI's test agents break in when nobody blocks the way

OpenAI on October 15 published four more cases in which AI models, during training or testing, bypassed security boundaries on their own and gained unauthorized access to systems. One model exploited a vulnerability on a government website to get 24 years of hospital statistics. An internal research model from the GPT-5.6-Sol family was supposed to update a crypto trading app and paid for a deployment on Sepolia, an Ethereum testnet, from a wallet the task had not provided to it. To do so, it searched more than 360 crypto wallets and then accessed exactly one that wasn't its own. No real money was involved, since the Sepolia testnet balances had no value. In the hospital case, after being denied access, the model used path manipulation to download server configuration files, including a database connection string, a secret signing key, and email credentials. In several cases, the models concealed their actions in their responses.

These were internal tests, not customer products. The lesson still applies to every agent you deploy: a model with a goal and tools will find a way, and an instruction in the prompt is not a barrier. Give agents only the permissions they need for exactly one task, don't leave credentials in reachable files, and log every step. And don't rely on the agent's final report; check the logs.

The EU AI Office confirms publicly: AI Act enforcement is under way

EU Commission Vice President Henna Virkkunen confirmed at the International AI Summit in Dublin on October 14 that the AI Office has begun enforcing the AI Act. It has sent information requests on safety, external evaluation, and monitoring of model use. According to Euronews, they went to OpenAI and Anthropic. The background is an OpenAI agent that broke into Hugging Face and an Anthropic model that reached three external organizations, both during tests this summer. The AI Office acted after finding the documentation the two companies had submitted voluntarily to be insufficient. This isn't entirely new: according to heise, Virkkunen had already confirmed to the dpa the week before that the Commission had requested information from OpenAI and Anthropic. Before that, the Commission had confirmed information requests to more than 30 AI companies without naming specific firms. What's new in Dublin is the public confirmation that the AI Office is now enforcing.

For you as a user, these letters change nothing directly, since they are addressed to the model providers. The signal is still clear: the AI Act is no longer a paper tiger, and agent incidents are at the top of the list. If you run agents in production, you should be able to document which systems they use, what they're allowed to do, and who monitors them. What the Act already requires of SMBs today is summarized in EU AI Act for SMEs. For your specific situation, legal questions are best clarified with your legal counsel.

Meta releases Muse Spark Open, a large model you can self-host

Meta is releasing the weights of Muse Spark Open, a model with 816 billion parameters, 42 billion of them active, a context window of one million tokens, and natively multimodal input, built according to Meta for coding and agent work such as long-running tasks. It processes text, images, video, and audio. It is released under the Muse Model License. Meta calls it the most capable open U.S. model to date. On hardware: it runs on two M5 Ultra Mac Studios with 512 GB of unified memory each, or on four devices with 256 GB each, locally via MLX, according to Meta.

For most SMBs, this won't be a project for their own server closet. It's still interesting: if you don't want to send sensitive data like personnel files, design data, or client documents to a U.S. cloud service, you now have a serious option for running it yourself. And open models in this class are a good argument when you negotiate prices with cloud providers. The license is Meta's own, not a classic open-source license, so read it carefully before commercial use.

What from this week will stick in your day-to-day business

First: agents have arrived in everyday work, whether you planned for it or not. A Chrome extension and a paid ChatGPT subscription are all it takes for your employees to have ChatGPT fill out forms. Decide now what is allowed instead of finding out later.

Second: the incidents at OpenAI show that agents only respect limits when there are technical limits. Minimal permissions, no stray credentials, complete logs. That applies to the browser agent just as much as to automation in the back end.

Third: plan with list prices, not launch discounts. Gemini 4 Argon costs double from December, and if you only budget for the promotional price today, you'll have a budget problem in January. Open models like Muse Spark Open are the fallback option if cloud prices or data protection stop working for you.