Mountain View, 28 October. Google has unveiled the Gemini 4 Carbon language model, which it calls its strongest model, and has so far opened access only to a subset of cybersecurity specialists; developers, businesses and ordinary users will get it later, Koray Kavukcuoglu, senior vice president of Google DeepMind, wrote on the company’s blog.
Carbon arrives two weeks after Gemini 4 Argon and, as Google describes it, is designed for long multi-step tasks: inside the company, according to the developers, it worked for days on complex engineering problems such as finding the cause of and fixing subtle bugs in large codebases. The main emphasis is on cybersecurity: the model scored 88.2% on SEC-bench Pro, a test of finding real vulnerabilities, 54.8% on ExploitGym, which checks whether a vulnerability can be exploited, and 71% on CWE-bench, a test of fixing vulnerabilities. According to Google, the Chrome browser team reported that Carbon was more than ten times better at producing high-quality patches than “another lab’s large frontier model”, which was not named.
In the results table Google published, Carbon is compared with its predecessor Argon, with OpenAI’s GPT-6 Astra and with Anthropic’s models Claude Opus 5.5 and Claude Fable 5.1. Carbon leads in most of the 17 tests. On Terminal-Bench 4.0, where the model works in the command line, it scored 65.2% against 59.6% for Claude Opus 5.5 and 59.1% for GPT-6 Astra; on the overall Artificial Analysis index it scored 59 against 58 for Claude Opus 5.5 and 57 for GPT-6 Astra. But not everywhere: on the FrontierSWE 2 coding test Carbon (61.2%) trails both GPT-6 Astra (65.5%) and Claude Opus 5.5 (62.3%), and on OSWorld-2.1, where the model operates a computer, it trails Claude Opus 5.5 (81.5% against 87.2%). All the figures come from Google itself; there are no independent checks yet.
The model launches at the same introductory price as Argon: $2 per 1 million input tokens and $10 per 1 million output tokens; a token is a fragment of text in which models measure their workload. The price is temporary: after the introductory period it will double to $4 and $20. Cyber defence specialists will get Carbon first, and access will then be extended in stages to trusted partners through the Fairwind programme. Google writes that it takes part in the voluntary process under which the US government tests models before release, and that it will keep refining safeguards before an open launch.
What it means: Google has released a model that, by its own tests, outperforms rivals at finding and fixing vulnerabilities, but is not giving it to everyone at once: defenders get it first so that hacking capabilities do not reach attackers earlier. The claim that Carbon is the best is true for most of Google’s tests, but not all of them.
Who benefits: Cyber defence specialists gain early access to a tool for finding and fixing vulnerabilities. Google gains a chance to strengthen its position in the frontier-model race with OpenAI and Anthropic. Developers will have to wait: no date for a broad launch has been given.
Background: Google DeepMind is Google’s artificial intelligence research division; it has been releasing the Gemini family of models since 2023. Anthropic, whose models Google cites for comparison, filed for an IPO in October.
We have also reported that Anthropic grew its revenue 14-fold ahead of its IPO.
