Contents7
Every month, we pick the AI news that changes something for a company of 20 to 500 people, each with a “what it changes” line. First edition: October 2026, closed on the 19th. Product announcements with no concrete effect on your choices are left out, and no model is ranked above another.
This format belongs to our Business AI section. To know where to start with AI, the starting point remains our guide to integrating AI in a business.
In short
Two large models are released, by Mistral and Google, and a well-known software company changes its default provider: the ability to switch models is becoming a buying criterion.
Retailers and large brands are starting to filter AI agents, and Meta rewrites its terms on crawling websites.
In France, a study by the Banque de France and HEC Paris measures a drop in permanent hiring of young people in the most exposed occupations, without making AI the only cause.
The law that is meant to transpose NIS2 in France is again withdrawn from the National Assembly’s agenda.
The month in dates
- 5 October
Mistral presents Large 4.
- 6 October
TechCrunch lists websites that block AI agents; France’s National Assembly withdraws the law transposing NIS2 from its agenda.
- 7 October
Google opens its SynthID Detector watermark checker to everyone.
- 12 October
Banque de France and HEC Paris study on hiring young people.
- 13 October
Gemini 4 Argon reaches developers in Google’s API.
- 15 to 18 October
The clause in Meta’s terms on crawling websites is debated.
- 16 October
Replit’s CEO says OpenAI is becoming its default provider.
- 19 October
The United States orders DeepSeek removed from the systems of federal civilian agencies.
Two new large models, and a provider replaced
Mistral presented Large 4 on 5 October, a “mixture of experts” model with more than 1,000 billion parameters, trained in Europe, first available as a preview through its API, with the weights announced for the end of October, according to Le Monde Informatique. Google, after a presentation reserved for cybersecurity partners on 30 September, opened Gemini 4 Argon on 13 October to developers on the paid tier of its API, according to its blog.
The same month, the CEO of Replit, an AI-assisted development platform, announced that OpenAI was becoming the default provider for its agent, for reasons of “cost, partnership, and performance”, with other models still available to users, TechCrunch reports.
What it changes: the best model for your tasks changes faster than your contracts. What protects you is being able to switch without rewriting everything: your own test sets, a layer that separates your application from the provider, and routing by task. The method is covered in evaluating an AI model before production and the AI bill, from tokens to routing.
From the field
On Palmora Property, a six-language real estate portal we operate, moving from one provider to another divided the cost by 60, and 95% of segments now come out of a translation memory without calling the AI again. The switch was only possible because the pipeline did not depend on a single provider.
Websites start sorting AI agents
An AI agent that books or buys for a user no longer gets in everywhere. TechCrunch cites Amazon, which has blocked Meta’s Muse agent on its store since September, Yelp, which only admits non-human traffic under its data licensing programme, and Delta, which has no partnership allowing a third-party agent to book its flights.
In the other direction, Meta published commercial terms that let it crawl websites linked to business accounts to train its products; the press reported the sentence removed on 17 October, without the preview read the next day confirming it. We detail what to check in can Meta train its AI on your website?
What it changes: for an online shop, the question is no longer “block robots or not”, but which actions an agent may take (read, log in, pay), and which you open to whom.
A French study on hiring young people
The Banque de France and HEC Paris’s LIFT lab compared, from January 2019 to March 2026, hirings and contract terminations according to how exposed occupations are to generative AI, Le Monde Informatique reports:
permanent hires of 21 to 29 year-olds in the most exposed occupations, from November 2022 to March 2026
in the least exposed occupations, over the same period
fewer permanent contracts a year for young people in the most exposed occupations, according to the authors’ estimate
The authors do not conclude that these occupations are disappearing. The IT sector was coming out of a hiring peak in 2021 and 2022, then went through rising interest rates and falling funding: AI “cannot be held solely responsible for the observed decline”. They also find no clear difference in total headcount between companies that use AI intensively and the others.
What it changes: the first effect shows up in entry-level hiring, not in headcount. A company that automates entry-level tasks has to ask how it will train its future seniors, the subject of our article on AI and young people’s first jobs.
What is moving on the rules side
Three facts of the month, plus a deadline:
- NIS2 postponed in France. On 6 October, the National Assembly withdrew the Resilience bill, which is meant to transpose the directive, from its agenda. What you can do without waiting for the law: NIS2 in France, what to do in the meantime.
- DeepSeek banned from US federal agencies. The Department of Homeland Security orders DeepSeek products, and third-party services built on them, removed from the systems of federal civilian agencies, according to Next. The decision does not apply in Europe, but it is a reminder to document where your data goes, whatever the provider.
- A watermark detector open to all. Google opened SynthID Detector to the public on 7 October; it recognises the mark of Google and its partners, not that of other tools, Le Monde Informatique notes.
- The deadline to note. On 2 December 2026, tools that generate images, video or sound must mark them: see marking AI-generated content, and the full timetable revised by the Digital Omnibus in the AI Act.
This is not legal advice: for a specific case, your DPO or a lawyer has the final say.
What a business leader can take from it
None of these facts requires action this week. Together, they point to three habits: keep the freedom to switch models, decide who may read and use your website, and follow a regulatory timetable that is becoming clearer. If you want to know what your current uses cost and what could move to another provider, our AI at the right cost offer starts from your actual bill.
Frequently asked questions
Which AI models were released in October 2026?
Among the announcements that matter for businesses: Mistral Large 4, presented on 5 October as an API preview with weights announced for the end of the month, and Gemini 4 Argon, opened on 13 October to developers on the paid tier of Google’s API. We do not rank them: only a test on your own tasks tells which one suits your use.
Is AI reducing the hiring of young people in France?
A study by the Banque de France and HEC Paris observes a sharper drop in permanent hiring of young people in the occupations most exposed to generative AI. The authors refuse to make it the only cause, pointing to the IT sector’s downturn after 2022, and see no difference in total headcount between companies that use AI intensively and the others.
Should a business stop using DeepSeek?
The US decision of 19 October only covers federal civilian agencies in the United States. For a European business, the right question is the same for every provider: where the data goes, under which contract, and which data must never leave your servers. A written grid by sensitivity level settles that choice model by model.
How can you avoid depending on a single AI provider?
By keeping your own test sets, a software layer that isolates your application from the provider, and routing that sends each task to the right model. The day a provider changes its price or its quality, you test the alternative on your real cases and switch without rewriting the application.
Sources
- Le Monde Informatique (new tab)lemondeinformatique.fr
- According to its blog (new tab)blog.google
- TechCrunch reports (new tab)techcrunch.com
- TechCrunch (new tab)techcrunch.com
- Le Monde Informatique reports (new tab)lemondeinformatique.fr
- According to Next (new tab)next.ink
- Le Monde Informatique notes (new tab)lemondeinformatique.fr
Jim · founder of Brain Plus AI
Also in business ai
Understand7 min
Can Meta train its AI on your website?
Meta's commercial terms that apply from 30 October 2026 let it crawl the websites linked to a business account and use their content to train its products. What was published, what may have changed on 17 October, and what you can check before the date.
Practical guide10 min
The cost of AI in business: tokens, routing and forecasting
What you really pay when you use an AI through an API: input and output tokens, the context sent back with every call, an agent's turns. And a method to forecast the bill before you receive it.
Stepping back8 min
Will AI take young people's first jobs?
No disruption is visible across the labour market as a whole, but a signal appears among young people, in the occupations where AI replaces the task. What the studies say, our view, and the real question for an SME: how to train its juniors tomorrow.