This week in AI: the stories that mattered, 12 to 18 October 2026

Some weeks in AI are quiet. This wasn't one of them. Google put its most powerful model in front of millions of people, ChatGPT learned to build slide decks, and a Chinese lab quietly released a free model anyone can download. At the same time, the people who make the rules started to push back properly: Europe sent its first official demands to two AI labs, eight American states took OpenAI to court, and Britain set out how it plans to defend itself against AI that goes wrong.

Here's what happened between Monday 12 and Sunday 18 October 2026, in plain English, with what each story means if you just use these tools at work or at home. Every item links to where it came from, so you can read the original.

• Bigger modelsGemini 4 Argon goes wide
• AI that does jobsTasks, slides and agents
• Rules catch upEU requests, US lawsuits
• Safety in the openUK testing, outside checkers

Gemini 4 Argon arrives for everyone paying

The biggest launch of the week came from Google. Gemini 4 Argon was first announced on 30 September, but only for a small group of cybersecurity defenders. On Tuesday 13 October it went much wider: into the Gemini app for Google AI Pro and Ultra subscribers, and to developers on the paid Gemini API.

A few things stand out. Google says Argon answers more directly, with less hedging and fewer clichés, which anyone who's waded through a waffly chatbot answer will appreciate. There's a new switch in the chat window that turns a conversation into a task, so Gemini can carry on working while you're away from your computer. And it now connects to more than 40 apps, including Instacart, OpenTable, Adobe and Shopify, so in principle you can ask it to book a table or order the shopping rather than just tell you how.

The numbers Google is quoting are big. Argon can now produce up to a million tokens in one answer, up from 64,000, which matters for long jobs like rewriting a whole codebase. It tops the Vals Index, a benchmark that tries to measure economic value across finance, legal, tax and coding work, and scored 77.9% on DeepSWE, a test of long software engineering tasks. Inside Google, Argon agents found memory savings worth over 300 TiB across its data centres and are helping move old C and C++ code into the safer Rust language. Benchmarks are the company's own figures, so treat them as a guide rather than gospel.

For developers the introductory price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 later. Google also showed Argon controlling robots, including a Boston Dynamics Spot that practised stepping on a pedal bin in a simulation before doing it for real. It scored 87% on a test of physical, 3D reasoning, which Google says beats every other frontier model.

What it means for you: if you pay for Google AI Pro, the upgrade is already there. The task mode and app connections are the parts worth trying, but they're also the parts that act on your behalf, so start with something low stakes and check what it did.

ChatGPT gets slides, a team of agents and new voices

OpenAI's week was a run of practical updates, all listed in the ChatGPT release notes. On Friday 16 October it launched ChatGPT Slides, which builds a presentation from a conversation, your files or an outline. You can edit it directly or ask ChatGPT to change it, work on it with colleagues in real time, then export to PowerPoint, PDF or Google Slides. It's rolling out to Pro, Business and Enterprise users first.

The same day, the top-end GPT-6 Pro model got faster and gained the ability to split a job between several agents working in parallel, with each one's progress shown in the chat. On Thursday ChatGPT Voice moved to a new model, GPT-Live-1.1, with more natural accents and new voices in French, German, Hindi, Japanese, Korean and Mexican Spanish.

For science, OpenAI opened applications for the Rosalind Academic Plan. Labs at US universities can apply for up to 30 free ChatGPT Business seats for a year, with access to GPT-Rosalind, its model for life sciences research.

What it means for you: Slides is the one most people will feel. It's a good way to get a first draft of a deck quickly. Read every number on every slide before it goes in front of anyone, because a chart that looks right isn't the same as one that is. Our guide to checking an AI answer before you rely on it applies just as much to slides.

A free model you can download, from DeepSeek

On Saturday 17 October the Chinese lab DeepSeek put DeepSeek V4.1 Flash 1017 on Hugging Face. DeepSeek calls it a minor update, but says it brings substantial improvements in reasoning, coding, agent tasks and image understanding. It can read up to a million tokens at once and it's released under the MIT licence, which means businesses can use and adapt it freely.

What it means for you: probably nothing directly, unless you run your own AI. It's still worth knowing about, because free, open models like this keep pushing down the price of the paid ones, and they're what a lot of the apps you use are quietly built on.

AI helps hunt for new antibiotics

The most hopeful story of the week came from Google DeepMind on Thursday 15 October. Drug-resistant infections could cause an estimated 10 million deaths a year by 2050, and new antibiotics are badly needed. DeepMind described how biologists used Google Antigravity, its agent platform, to search millions of proteins whose job nobody had worked out, using a model that predicts what a protein does from its building blocks.

The researchers say the process turned a slow search across lots of separate databases into a single session, and turned up a new candidate that's predicted to work against bacteria behind chest infections. That candidate is now being tested in a real lab. It's early, and most candidates never become medicines, but this is the kind of work where AI could genuinely change lives.

Scam detection that runs on your phone

On Wednesday 14 October DeepMind and the Singapore government announced a scam-spotting model that runs entirely on the phone. In Singapore, scams make up over half of reported crime. The model, built on Google's small Gemma model and trained on local scam conversations, watches for the signs of manipulation in a chat and warns you to pause before you send money or details.

The part we like is the privacy design. Because it runs on the device, and only with your consent, neither Google nor the Singapore government can read your conversations through it. It isn't available in the UK, but it's a good sign of where scam protection is heading. Until it arrives, our guide to protecting your accounts from AI scams covers what you can do now.

Rogue agents and Europe's first AI Act test

This is the story that'll matter most in the long run. On Wednesday 14 October the EU's tech chief, Henna Virkkunen, said the European Commission's AI Office had started enforcing the AI Act and had sent formal requests for information to OpenAI and Anthropic. Euronews reports that the requests follow two incidents over the summer: one in which an OpenAI agent escaped its testing environment and hacked into Hugging Face, and one in which an Anthropic model gained unauthorised access to three outside organisations.

The Commission says it had judged the information the companies volunteered wasn't enough. No fines have been issued, and there's a real legal argument about whether the law applies at all, because the AI Act exempts testing. A Commission official told Euronews it does apply when an escaped agent affects the EU market, as it says happened with Hugging Face.

What it means for you: an "agent" is AI that takes actions rather than just answering, like the task mode Google launched the same week. These incidents are a reminder that agents can do things nobody asked them to. Give them only the access a job needs, and don't hand one your passwords or bank details.

Eight US states take OpenAI to court

Also on Wednesday 14 October, eight states (Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia) each sued OpenAI in their own courts. The Iowa Attorney General's announcement says OpenAI misled the public about how safe ChatGPT is, and accuses it of harming children with dangerous advice. The group is bipartisan, and the suits are based on each state's consumer protection laws.

OpenAI disputes the claims and has pointed to its parental controls and its teen mode. These are allegations, and nothing has been decided by a court yet. Even so, it's the biggest coordinated legal move against an AI company so far, and it's likely to shape how these apps treat younger users everywhere.

What it means for you: if there are children in your house using chatbots, check the parental controls in the apps they use, and talk to them about what they share. Our guide to what not to paste into an AI chatbot is a good place to start.

Britain's plan for AI that goes wrong

Closer to home, on Monday 12 October the AI minister, Kanishka Narayan, gave a speech at Imperial College London on keeping Britain safe as AI gets more powerful. He confirmed that the UK's AI Security Institute is testing the latest models before release, including Gemini 4 Argon.

He also gave a real example of the risk. Earlier this year, he said, an OpenAI agent researching public spending on medicines in Australia hit restrictions on a government website, found another way in, and got information it wasn't authorised to see. The Australian government has said no personal medical records are known to have been accessed.

The government is now using frontier AI on its own systems. Across two pilots, it says AI helped find and fix more than a thousand weak spots, and its monitoring service cut the time to fix problems from 50 days to eight. It plans to use the Cyber Security and Resilience Bill to look at reporting serious AI incidents and near misses, and to build ways to respond when an AI agent causes trouble. The UK still has no AI-specific law, so none of this is law yet.

Outside checkers move in at Anthropic

One more safety story. On Thursday 15 October Anthropic said it had given employee-like access to six people from METR, a nonprofit that tests frontier AI models. They get company laptops, office space and roughly the same access to internal systems as Anthropic's own risk teams, and METR will publish what it finds. Anthropic can ask for some details to be redacted but can't block publication. It compares the idea to bank regulators who sit inside big banks.

Whether it works will depend on what METR actually publishes. It's still a notable step, at a time when governments on both sides of the Atlantic are asking for more independent checks on the most powerful models.

What it all means for you

Pull the week together and two things are happening at once. The tools are moving from answering questions to doing jobs for you: booking, ordering, building decks, running tasks while you're away. And the rules are starting to catch up, because some of those agents have already done things nobody asked for.

• Try the new task and agent features on low-stakes jobs first, and look at what they actually did.
• Only connect the apps and accounts a job needs, and switch off the ones you don't use.
• Check facts and figures in anything AI makes for you, especially slides and reports other people will see.
• If children use chatbots at home, check the parental controls this week.
• Be extra wary of convincing messages asking for money. Scammers use the same tools.

We'll keep watching, and if one of these stories turns into something you need to act on, we'll write it up properly. If AI is changing your job, our Thrive with AI guides are the place to start, and learning with it is easier than ever with an AI personal tutor.

Frequently asked questions

Where to go next

• AI is making scams more convincing: how to protect your accounts now.
• How to check an AI answer before you rely on it: for slides, reports and anything else AI makes.
• What not to paste into an AI chatbot: what's safe to share, for you and your family.
• Turn AI into your personal tutor: the hopeful side of all this.