Gemini 4 Carbon Goes Public as AI Proves Signal's Code Is Secure

November 12, 2026 · Cortex Solutions

Google opened Gemini 4 Carbon to developers and businesses on Nov 10, and DeepMind used it to formally verify Signal's protocol code. What it means for Saudi firms.

On November 10, 2026, Google announced that its Gemini 4 Carbon model is now available to developers, enterprises, scientists and consumers, two weeks after a limited release to a group of cybersecurity specialists. On the same day, Google DeepMind published research saying that AI agents powered by the model had mathematically proven key security properties of the actual code behind the encryption protocol used by Signal. Together, the two announcements point in a clear direction: AI models are no longer only writing code, they are starting to be used to check it and prove it is safe.

From a limited launch to general availability

According to a Reuters report published by Khaleej Times on October 28, 2026, Google unveiled Gemini 4 Carbon two weeks after introducing Gemini 4 Argon, saying it delivers advanced performance on cybersecurity tasks such as vulnerability discovery, exploitation and software patching, and that it was being made available first to a subset of cyber defenders through the Fairwind Program while safeguards were tested.

In Google's official post dated November 10, Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google, described Carbon as the company's "strongest model", with frontier performance in cybersecurity and in long-horizon agentic workflows across complex knowledge work, software engineering and scientific reasoning.

Google explains that Argon remains the right choice for most software engineering and knowledge work, while Carbon is meant for harder tasks such as complex multi-step coding and debugging and open-ended scientific research. It uses more tokens than Argon because it explores, reasons and runs verification loops before answering.

Where is it available and what does it cost?

According to Google's post, the model is available through:

• The Gemini API on the paid tier in Google AI Studio and Android Studio.
• Gemini for business for enterprises, via the API or the model picker, as well as Google Antigravity and CodeMender.
• The Gemini app and AI Mode in Search, for Google AI Ultra subscribers only.
• The Fairwind Program for cyber defenders, and a free six-month access program for scientists, subject to eligibility, capacity and selection.

On pricing, Google says it is an introductory price matching Argon: $2 per million input tokens and $10 per million output tokens. A footnote in the post says the introductory period ends on November 30, 2026, after which the price becomes $4 for input and $20 for output per million tokens. The post says nothing about availability by country or region.

What Google says about performance and safety

Google lists several results, all of them the company's own figures:

• The model ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than 2X lower cost per task than the next best external model.
• On an "agentic honesty" benchmark of 375 tasks, Carbon passed 92% of tasks, compared with 70% for Argon.
• In safety testing, the model blocked 99% of attack trajectories, and the per-attempt breach rate fell to 0.8%, down from 3.1% on Argon.
• Trajectory Labs tested the model externally and found zero new universal jailbreaks across 132 hours and more than 13,000 requests.
• Google says Wiz's Scan for Good program used the model to scan two major SaaS providers and uncovered zero-day vulnerabilities in under two hours.

Mathematically proving the safety of Signal's code

The most striking news came in a Google DeepMind post dated November 10, by Pushmeet Kohli and Swarat Chaudhuri, which introduced a framework called AlphaProof Code. It is an agentic framework that combines human expertise with AI agents: humans decide what needs to be proven, the agents write the proofs, and the Lean language checks every proof automatically.

According to the post:

• The team proved key security properties of the actual code of the Signal protocol, not only its theoretical design, including keeping messages protected before and after a device is temporarily compromised.
• The agents produced more than six million lines of Lean, and only 1% of the codebase needed human review.
• The verification took four weeks, a task DeepMind says historically took expert teams years.
• The team reintroduced two real, previously fixed bugs into an isolated version of the code, and in both cases the system produced a machine-checked proof that the buggy version was insecure.
• DeepMind describes the work as the first formal verification of a widely deployed cryptographic protocol implementation at this depth, and says it is open-sourcing its verification artifacts.

The post quotes Ehren Kret, Chief Technology Officer of Signal, as saying that using AI agents to translate live code into valid Lean theorems is "a major breakthrough." These results are published by Google itself, and we have not yet found an independent published assessment of them.

What this means for Saudi businesses

• The introductory price window is short: anyone who wants to test Carbon on real use cases can use the current price until November 30, while budgeting on the doubled price after that, since the model naturally uses more tokens.
• Not every task needs the most powerful model: Google's own guidance is that Argon is enough for most work and Carbon is for long, complex tasks. Choosing the model per task keeps the budget under control.
• Cybersecurity is entering a new phase: if models can find vulnerabilities this quickly, attackers will try the same. Reviewing and updating your systems and scanning your code regularly is no longer a luxury.
• Verify before you trust: the published figures come from the developer, so it is best to test the model on your own data and tasks before any decision, and to comply with the Personal Data Protection Law when sending any data to external services.

At Cortex, we help companies build AI agents and automate their work in a practical, well-planned way. Learn about our service: https://cortex.sa/services/ai-automation