Google has a new frontier model, and this time it is not aimed at everyone first. On October 28, 2026, Google announced Gemini 4 Carbon, describing it as its strongest model so far, with state-of-the-art cybersecurity and long-horizon capabilities. It arrives just two weeks after Gemini 4 Argon, and its first users will be a small group of cyber defenders — not the general public.
That staged rollout is the real story. Carbon is built for the hard, long-running work behind secure software: finding flaws, checking whether they can actually be exploited, and fixing them at the root cause. Google says its internal teams have watched Carbon run for days on multi-step engineering jobs, such as tracking down subtle bugs across large codebases, persisting through dead ends instead of giving up early.
What Google claims Carbon can do
Google's announcement, written by Google DeepMind chief technology officer Koray Kavukcuoglu, puts cybersecurity at the centre. On its own reported benchmarks, Carbon scores 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. Google also says Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59.
These are company-reported figures, and they should be read as claims to be tested by independent researchers, not as settled verdicts. Benchmarks measure specific tasks under specific conditions; they do not mean a model will find every flaw in a real production codebase, or that it should patch live systems without human review. Still, the direction is clear: Google is pitching Carbon as a defence tool first, a general chatbot second.
The same defence-first theme runs through other recent security work, including the UK's big bet on AI security covered in our report on the world's first AI biosecurity lab.
Better coding, fewer made-up claims
Beyond security, Google says Carbon improves on Argon across coding and agentic tasks. Based on user feedback from the previous release, it claims an 89-point jump on the WebDev Arena leaderboard, to fourth place, for full-stack web development, alongside gains in scientific tasks and multimodal understanding.
Google also highlights reliability. In internal testing, Carbon cut unsupported claims by about half compared with Argon, stating methodological caveats directly rather than fabricating missing evidence. It also says Carbon retrieves relevant sources and surfaces discrepancies when facing ambiguity, while reducing task overreach by 26% and under-service by 29% during multi-step workflows. For long agent jobs, that kind of restraint matters as much as raw capability — an agent that does too much, or too little, can be as unhelpful as one that gets the answer wrong.
That focus on agentic coding echoes the wider shift in developer tools this week, from Google Cloud's Gemini agent for work to GitHub's new Copilot Fusion system, which picks and orchestrates models for coding tasks instead of asking developers to choose one themselves.
Who gets Carbon first — and what it will cost
Google says it is taking a deliberate approach to rollout and is actively engaged in the U.S. government's voluntary process for pre-release model testing. Starting this week, Carbon goes to a subset of cyber defenders, then gradually to additional trusted partners through Google's Fairwind Program. Developers, businesses and consumers will get access, Google says, as soon as possible — but no date has been given.
When it does reach general API access, Google says Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input price. That pricing describes future commercial access, not a service anyone can sign up for today.
The limited start also explains why Google's own Chrome team finding — that Carbon produces high-quality patches more than 10 times better than another lab's large frontier model in its internal tests — needs outside confirmation. Until more defenders and researchers can run Carbon themselves, the strongest evidence is Google's word plus its published benchmark table.
Why a cyber-first frontier model matters
Most frontier launches are framed as general intelligence getting smarter. Carbon is framed differently: as infrastructure for defence. If a model can reliably find, validate and patch vulnerabilities, it could change the economics of defensive security work, where skilled reviewers are scarce and codebases keep growing. The same capability in the wrong hands, or with the wrong permissions, could also amplify risk — which is why Google stresses safeguards, pre-release testing and a phased rollout before wider access.
Google's recent AI search expansion shows how quickly its products move from announcement to worldwide rollout, as seen in its AI Mode ads reaching 10 new countries. Carbon is moving the other way: slowly, and to a trusted few first. That caution may frustrate users who want to try it today, but for a model built to touch security-critical code, caution is the point.
For now, the honest summary is this: Google has announced its strongest model yet, led with cybersecurity, and asked the security community to test it first. Whether Carbon lives up to its benchmark claims will be decided not by launch-day headlines, but by what those defenders find when they actually use it.
What do you think — should powerful cyber-defence models be rolled out to trusted defenders first, or should everyone get access at the same time? Tell us in the comments below.
Comments
Post a Comment