Europe's biggest open model, and a free year on Claude
Edition of 2026-10-07
# AI For All ### Wednesday, October 7, 2026
As of today, the technology lead at a European company can try a frontier-class model that was trained and runs inside the EU, a third option beside the American closed models and the Chinese open ones.
The model Europe built to do the work others refuse
Mistral opened a public preview of Mistral Large 4 on Tuesday, and calls it its strongest model yet. It was trained from scratch in Mistral's own European datacenters, and you can try the preview today on Mistral Studio. The open weights, the part that turns a demo into an alternative, are promised by the end of the month.
The detail worth reading twice is the security pitch. Mistral says the model scores 82% on an independent test that asks it to reproduce a real vulnerability and then patch it, the highest of any model. Several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on that test because they refuse the task at all. The company's argument is uncomfortable and hard to dismiss: the same refusal that stops an attacker also stops the defender.
If you run security operations, that is the trade on the table. Capability with fewer refusals, on a preview you cannot yet download. Mistral also reports 93% of the challenges in Cybench, and training data spanning more than 160 languages, which is the sovereignty case in one line.
A free year on Claude, if you can get in
Anthropic expanded its Claude for Startups program on Tuesday. Approved founders now get a free year of Claude Team, its paid plan for groups, plus $1,000 in API credits to build with. The same offer lists up to $45,000 in value from partner companies, along with office hours with Anthropic's applied AI team.
The two costs that stop a small team from standardising on one model are the seat price and the metered bill. This offer covers both for a year, which matters more to a five-person company than to a funded one. Read the terms before you plan around it: access runs through an application, and approval is not automatic.
If you are picking a model to build on, a year of free seats is a real nudge. It is also a lock-in with a one-year fuse, and that is worth noticing while the clock runs.
Anthropic turns the filters down, for the defenders it can vet
The same company moved on the other side of the same problem. Anthropic relaunched its Cyber Verification Program as three access tiers, each with its own vetting and its own level of safety blocks. The top tier, reserved for a small set of verified organisations, keeps the fewest. The everyday tier is open to security teams at companies, nonprofits, hospitals and utilities, and Anthropic says it aims to answer those applications within a few days.
The number it leans on is an outcome, not a feature. Partners in its earlier Glasswing program uncovered at least 129,000 verified software vulnerabilities between April and July, and its own scanning added 5,500 more. In Anthropic's own testing, 46 of the 50 trials were blocked in the middle tier, which is the filter the tiers are built to tune.
For a small security team, the question is not whether the model can do the work. It is whether the model will, on the day the work is legitimate and the request still looks like an attack. Anthropic is now selling the answer as a service, behind vetting.
Search that never leaves the device
Google released EmbeddingGemma 2, an open model that sorts text, images, audio and video into one searchable space. The licence is commercially permissive, and Google says the smallest version runs inside a phone's memory, on a Pixel. This is a building block, not an app, and it is built for work that never touches a server.
The claim to check is the convenience. Google says the model can find a video clip from a voice note, or search hours of audio from a typed query, without the files leaving the device. The first EmbeddingGemma passed 20 million downloads, which is why the sequel matters more than its size suggests. Developers can also cut storage needs by up to 6x, and the model can hold 5.5 minutes of audio at once.
For anyone whose company sits on recordings it would never upload, this is the useful shape of a release: the search comes to the data, instead of the data going to the search. The reservation is the same with every small model. Someone still has to build the tool around it, and that someone is not paid by this newsletter.
Weak signal
OpenAI published a catalogue of mathematical manuscripts, grouped into families, on a public repository. The material came from an unreleased internal model, and many of the results carry machine-checkable proofs, with more promised as they are formalised. Mathematics is the useful test case, because a wrong answer is caught rather than argued, and the method, a few hours of thinking compute per result, is the shape of research work to come. Nothing here is a tool a reader can pick up today, and that is exactly why it is a signal rather than a story.
The pattern this week is access. Mistral opens a preview and holds the weights, Anthropic opens a door and keeps the vetting, Google opens the model and lets the work stay local. What you can do today is the question every one of them is answering a different way.