Google Unveils Gemini 4 Carbon: A Security-Focused AI Model
· Cortex Solutions
On Oct 28 Google unveiled Gemini 4 Carbon, built for finding and fixing security flaws. Its benchmarks, analysts' caveats, pricing, and what it means for Saudi firms.
On October 28, 2026, Google announced its new model, Gemini 4 Carbon, describing it on its official blog as "our strongest model" with advanced cybersecurity capabilities and the ability to handle long, multi-step tasks. The announcement came in a post by Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google, only two weeks after the launch of the previous model, Gemini 4 Argon. Notably, the model will not be available to everyone yet, but only to a limited group of cybersecurity defenders.
What does Gemini 4 Carbon offer?
According to Google's announcement, the model focuses on coding, cybersecurity and long-running work. The company says its internal teams have seen the model work for days, tracing the root causes of bugs and fixing them in large codebases without giving up at dead ends. It also reports improvements in coding, agentic tasks, science, web development and multimodal understanding.
On the security side, Google presents the model as covering the full chain: finding vulnerabilities, validating whether they can be exploited, and then fixing them. The figures the company published are:
- 88.2% on SEC-bench Pro for finding vulnerabilities in real-world software.
- 54.8% on ExploitGym for assessing whether vulnerabilities can be exploited.
- 71.0% on CWE-bench for patching vulnerabilities.
- First place on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59.
Help Net Security, in a report by Anamarija Pogorelec on October 29, said Google's Chrome team reported the model produced more than ten times as many high-quality fixes as a large model from another lab that was not named, and that Wiz, Google's cloud security platform, said it led web penetration testing benchmarks.
Safety: figures reported by the company
Google says the model uses the same safety stack that Argon was built on. It reports that unsupported claims were cut roughly in half compared with Argon, that high-severity alerts about behavior conflicting with intended instructions fell by almost 80% in internal coding environments, and that the success rate of indirect prompt-injection attacks on the Gray Swan benchmark was 0.0%, versus 0.7% for Argon.
What do analysts say?
VentureBeat's report, written by Carl Franzen on October 28, put these numbers in context. According to Google's own chart, Claude Opus 5.5 scored 58 and GPT-6 Astra scored 57 on the same index, so the margin is very narrow. Both rival models also lead Carbon on some software engineering and computer-use tests.
The report flagged important caveats:
- Google has not published a technical report or details about the model's architecture or size.
- The cybersecurity evaluations were run "without guardrails," so customer results may differ.
- On an independent factual-knowledge test by Artificial Analysis, correct answers rose from 48% to 58%, but the "hallucination" rate, as that test defines it, also rose from 4% to 11%.
- There is no public API for the model yet.
Help Net Security also noted that most of the figures come from Google or its partners, with no published independent verification.
Availability and pricing
Google says access starts this week for a group of cyber defenders, then expands gradually to trusted partners through the Fairwind Program, and that it is taking part in the US government's voluntary pre-release model testing process. Developers, businesses and consumers were promised access "soon," with no specific date.
The company announced an introductory price of $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens, rising after the introductory period to $4 and $20. Google has not said when that period ends, according to VentureBeat.
What does this mean for Saudi businesses?
- AI is now a cybersecurity tool on both sides: models that can find and fix vulnerabilities also mean attackers will find them faster. Updating systems and closing known vulnerabilities can no longer wait.
- Do not base your decision on lab numbers alone: the gaps between the leading models are small, and most figures come from the companies themselves. Test a model on your real tasks and data before committing.
- Watch the long-term running cost: the introductory price may double later. Budget on the final price, not the introductory one, and use caching where available.
- Keep human review in the loop: the higher hallucination rate on one test is a reminder that any model can be confidently wrong, especially with contracts, financial figures and sensitive data.
- Respect local regulations: when using external models with customer data, make sure you comply with the Personal Data Protection Law and the requirements of the relevant authorities in the Kingdom.
At Cortex Solutions, we help companies choose the right AI model and connect it to their systems securely and cost-effectively. Learn about our AI business automation services: https://cortex.sa/services/ai-automation
Sources: Google's official blog (October 28, 2026), VentureBeat (October 28, 2026) and Help Net Security (October 29, 2026).