Google has announced Gemini 4 Carbon. According to the company, access starts with selected cyber defence teams; more partners will follow gradually. The announcement gives no firm date for general availability to developers, businesses or individual users.
What this means for you: a model announcement is not yet a feature you can use in your app today. Performance claims and access are separate questions. We have not tested Carbon ourselves.
Artificial Analysis has moved its Intelligence Index to version 5. New and revised tests change its composition; according to the provider, 60 percent of the index now relies on private test sets or unpublished answers. A separate Progress Index is intended to make longer-term development comparable.
Our take: comparing old and new scores directly can confuse a change of yardstick with an improvement in the model. Check the version, tasks and settings — then look at your own task.
Anthropic is expanding its AI for Science programme and introducing 20 participating foundations, institutions and research groups. The programme supports scientific projects with Claude credits. The announcement also describes joint work on data, software and standards.
Why it matters: beyond new model names, it matters who can use these tools for concrete research. A partnership alone is not evidence of a discovery or a measured speed-up.
Week of 19–25 October 2026 · Sources checked 25 October 2026
01 · Research · 22 October
AI helps prove properties of software
Anthropic reports on mathematical proofs for selected properties of software. Such proofs only hold under precisely stated assumptions. It's an interesting step in how programs are checked, but it does not establish general freedom from bugs. For us, the real news is the line between “tested” and “proven under conditions”.
Limits
Research by the company itself, not reproduced by Hexpex. Article from 22 October, post on X on 23 October.
OpenAI describes a shielded training environment in which models work with a stored copy of the web. It responded to unwanted network actions during internal research. The change affects the training of new models. It does not mean that web search has been switched off for regular ChatGPT users. The report shows why a system's access and limits matter too.
Limits
The provider's own account, not independent proof of security. Affects internal training, not ChatGPT web search.
DigUp is meant to find files even when you no longer remember the file name. According to the developer, the app processes images, PDFs and audio, among other things, locally. It first downloads a model, and an optional update check goes online. It sounds useful, but we haven't tested it yet. The Windows version is labelled beta.
Limits
Developer claims; we have not installed or tested the app. Spotted on 24 October, not a verified release date. Privacy claims have not been independently checked by Hexpex.