Control: Mistral Large 4 sells sovereignty, OpenAI watermarks its text in Europe, Anthropic arms the defenders

Three stories about the same negotiation: who sets the rules for the most powerful technology on the planet. Europe wants its own champion. OpenAI has to sign its work, because Europe requires it. And Anthropic hands its most powerful models to professional software breakers. Three rooms, one question: control.

Signal one: the European champion is back

On Tuesday, October 6, French start-up Mistral unveiled Mistral Large 4, its new flagship model. Co-founder Guillaume Lample says it "narrows the gap" with the best models in the world. The announcement places it among the world's best open models, meaning freely accessible and modifiable.

The context is harsh

Mistral's last major release dated from December 2025, and its models had slipped to 24th place in Artificial Analysis's aggregate ranking, a benchmark comparison across a broad set of tests. At the top, the Americans: OpenAI, Anthropic, Google. Behind them, a Chinese pack: Moonshot, Alibaba, Z.ai. A French start-up in a race measured in hundreds of thousands of chips.

The numbers, preliminary

"Preliminary" is Mistral's word. On Deep SWE 1.1, a benchmark for long programming tasks, Large 4 scores 63%. That is the level of GLM 5.3 from China's Z.ai, twelfth overall. The best models reach 74%.

Mistral also claims to match the best open models in finance, cybersecurity and geospatial analysis, meaning reading satellite images to assess, for example, disaster damage. Airbus and BMW are customers.

Two details matter more than the scores

Large 4 goes fully open source on October 27, after safety testing. It will be the default model in Mistral's assistant, Vibe, though users will still be able to choose Chinese models such as GLM where they remain better. I note that product honesty: letting users pick a competitor when it is better is a rare move.

The sovereignty argument. Pierre Stock, Mistral's chief scientist, told Le Monde that open source AI "guarantees that access to the model can never be cut off." He was pointing at the White House, which had temporarily restricted access to Anthropic's Mythos and Fable models. An open model running on your own servers cannot be taken away from you for political reasons.

Let me be precise about "open" here, because the word is often vague. An open-weights model is one whose parameters are published: anyone can download it, run it on their own machines and adapt it. A company using it no longer depends on an API that a vendor, or a government, can shut.

Why it matters

The arithmetic of the race changes shape. Large 4 was trained on 4,000 chips, two to three times fewer than Chinese start-ups have, and nothing next to the hundreds of thousands of chips of the American leaders. Mistral raised 3 billion euros in early September and targets a gigawatt of data centers by 2030. The gap between 4,000 chips and the Chinese frontier is now counted in one or two months. Efficiency becomes a weapon in its own right.

Sovereignty becomes a product feature. When the White House can restrict which models you are allowed to use, a guarantee that nobody can cut your access is worth more than a few benchmark points.

For brands choosing an AI vendor, the question widens. There is the question of how intelligent the model is. There is now the question of whether it stays available when politics changes. Europe has just become a credible answer to that second question. For a European, Moroccan or African company building critical processes on AI, I would add this criterion to any selection grid today.

Signal two: the watermark nobody can check

On October 5, OpenAI announced it would embed an invisible watermark in text and code generated by ChatGPT and Codex in the European Union. The system is called textGrain.

The driver is the law. Article 50 of the EU AI Act requires providers to mark synthetic content in a machine-readable way. Those rules have applied since August 2, 2026. Companies already on the market have until December 2 to comply. Fines reach 15 million euros or 3% of global turnover.

How it works

The model subtly steers its word choices while writing, according to a secret key. It leaves a statistical pattern that a detector can spot and a human cannot see. The signal survives copy and paste.

Put simply: at each word, a language model chooses among several plausible candidates. The watermark slightly favors some candidates, designated by the key. Over a long text, those small preferences add up to a measurable signature. Over a short text, there are fewer words, so less signature.

OpenAI published a technical report, co-authored with researchers from the University of Pennsylvania and Yale, claiming textGrain matches or beats alternatives, including Google DeepMind's SynthID. The watermark does not identify the user, and OpenAI plans to open source the technique.

The fine print, published by OpenAI itself

• On short texts, detection is around 80%.
• Replace a tenth of the words with synonyms and accuracy drops from 92% to 66%. A motivated cheater only needs a thesaurus.
• Detection is weaker on math and on translated text.
• OpenAI's warning: the absence of a watermark does not prove human origin.

The detector in a locked room

The sharpest detail is the detector. OpenAI built it and, for now, gives it only to approved researchers and expert organizations. No timeline for wider access. A teacher in Lisbon or an editor in Warsaw facing a suspicious text cannot check anything today. A transparency tool the public cannot use.

The map tells its own story

Anthropic watermarked Claude's text from August, worldwide, and took criticism from users who felt they supplied the ideas while the model was only a tool. OpenAI chose the opposite: on by default where a regulator can fine it, off by default for developers everywhere else, offered as an option since October 5.

The Wall Street Journal reported in 2024 that OpenAI had already built a working watermark and shelved it, fearing users would move to rivals without one. The law removed that fear in Europe. Outside Europe, fear apparently still decides.

What changes

The question is who gets to tell the real from the synthetic. Today the answer is: almost nobody. A detection system reserved for approved researchers, which collapses in front of a thesaurus, is compliance infrastructure. Accountability infrastructure remains to be built.

For brands, the practical angle is this. Content provenance is becoming a legal category, with a hard deadline. If your marketing, reports or communications contain AI-generated text, someone in your chain will need to attest to it. I would recommend three habits: know which brand content goes through an AI, keep a record of it, and decide in advance what the brand declares publicly. The watermark is the first instrument. It is not reliable yet. I read it as an emerging standard, without treating it as a verdict.

Signal three: arming the defenders

On October 6, Reuters reported that Anthropic is expanding the program that lets vetted cybersecurity professionals test its most powerful models with fewer safeguards.

The reason is a striking number. Partners in Project Glasswing found at least 129,000 verified software vulnerabilities between April and July. Anthropic's own scanning found 5,500 more. More than 33,000 were rated critical or high severity. The company says the real figure is probably at least five times higher, because only a limited number of partners reported results.

The program

Called the Cyber Verification Program, it merges two schemes: Glasswing, which gave defenders of critical software access to Claude Mythos, and the original CVP, which gave vetted teams lighter safeguards on Claude Opus and Sonnet. Three tiers:

1. Defense: incident response and malware analysis. Open to security teams, infrastructure operators, open source maintainers and researchers with a disclosure track record.
2. Red Team: adds authorized penetration testing. Organizations only.
3. Specialized: the fewest restrictions, for a small group testing critical systems such as power grids, flight systems or interbank transfers.

All three tiers get Claude Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models. Every member is vetted jointly with the US government.

Six months earlier

In April, when Anthropic unveiled the Claude Mythos preview, the dominant fear was that AI would hack software faster than anyone could secure it. Six months later, the same company runs the largest vulnerability discovery operation ever reported and hands the capability to defenders, under government vetting.

What changes

Attack industrialized first. The question was always whether defense would get the same tools. The answer is arriving: yes, through a front desk, with three tiers and a government list.

For businesses, the message is direct. Your software has never been examined so closely. More than 33,000 critical or high severity vulnerabilities found by one company's partners in four months means your patching discipline is now measured against a machine that never sleeps. I connect this to Microsoft's report from the week before: flaws exploited in under 24 hours, patches applied in 30 to 60 days. The audit trail has become the only proof you were watching.

The thread of the day

Europe builds its champion and sells the guarantee that nobody can take it away. OpenAI signs its text and keeps the detector in a locked room. Anthropic arms the defenders, after the government has checked their names.

Three rooms, one negotiation: who controls the most powerful technology on the planet, and on what terms. The terms are being written now, in open weights, secret keys and vetting lists. Those who do not read them while they are written will inherit them.

The noise filtered. The signal kept.