Two announcements from 28 October that mainly affect developers and security teams. Google presents a new top model you cannot use yet, and GitHub now lets Copilot decide itself which model carries out a task.
Gemini 4 Carbon: strong at vulnerabilities, not yet for everyone
Google calls Gemini 4 Carbon its strongest model so far. The emphasis is on security: finding vulnerabilities in software, checking whether they can really be exploited, and writing a fix. Google reports scores of 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (demonstrating exploitation) and 71.0% on CWE-bench (writing patches). On the Artificial Analysis Intelligence Index v5.0, the model ranks 1 according to Google, with a score of 59.
You can't use it yet. This week only some of the "cyber defenders" get access through the Fairwind programme, and more partners follow step by step. Developers, businesses and consumers come "soon", with no date. The API price is already fixed: during the introductory period $2 per million input tokens and $10 per million output tokens, after that $4 and $20. Cached input gets a 95% discount.
What does this mean for you? For now, not much, unless you work for a security partner of Google. The announcement says nothing about EU hosting or data processing. Once Carbon is widely available, the same business terms will probably apply as for the other Claude models, but Google still has to confirm that. You can see how Claude currently scores on privacy on the Claude page and in the GDPR check for Claude in software development. If you want to compare it with the competition, look at Gemini vs Claude and ChatGPT vs Claude.
GitHub Copilot Fusion: effort level instead of model choice
At the Universe conference, GitHub made Copilot Fusion generally available across all Copilot plans. You no longer choose a model but an effort level: Standard for everyday work, High for harder tasks and Max for the most demanding work. Fusion then decides which models take part. That can be a single model, a cheap model that scales up when in doubt, or a setup in which one model writes a draft and another checks it. It works in VS Code, the GitHub Copilot app, the Copilot CLI and JetBrains.
You pay the standard rate per model used. On paid plans there is a temporary 30% discount. According to The New Stack, Fusion Max scores almost the same as Claude Code with Opus 5.5 on the Coding Agent Index from Artificial Analysis, at roughly half the API cost. Those figures come from GitHub itself.
What does this mean for you? If you work in a team, one point matters most: an administrator of an Enterprise account decides which models Fusion may use. If your organisation has not approved certain models, for example because of where the data goes, switch them off for Fusion too. Otherwise Copilot might choose a model you would never choose yourself. Check the GDPR check for GitHub Copilot, the GitHub Copilot page or the alternatives to GitHub Copilot.
In brief
- Gemini 4 Carbon has been announced and the price is fixed, but there is no date yet for wide use.
- Copilot Fusion chooses models itself. As an administrator, set which models are allowed.
- All models and their privacy scores are shown side by side under models and safety. If you are a developer, also look at AI for software development.
