Google has announced Gemini 4 Carbon, its strongest model yet, built for cybersecurity defense and long-running engineering work. Announced on October 28, 2026, just two weeks after Gemini 4 Argon, Carbon tops the updated Artificial Analysis Intelligence Index at 59 points, ahead of rival frontier models, and posts leading scores on vulnerability discovery, exploitation and patching benchmarks. Access starts this week with a small group of cyber defenders through Google’s Fairwind Program, with wider availability for developers, businesses and consumers promised soon but not yet dated.
Two Frontier Models In Two Weeks
Carbon builds directly on Argon, which Google launched two weeks ago, and the pace itself is part of the story. Google describes Carbon as illustrating how fast it is pushing the frontier forward. The company is also taking part in the United States government’s voluntary process for pre-release model testing, a deliberate rollout choice in contrast to putting a new model in everyone’s hands on day one.
That caution is not just presentation. Carbon is a model designed to find, prove and fix software flaws, which makes it powerful for defenders and, in the wrong hands, useful for attackers too. Limiting the first cohort to vetted cyber defenders lets Google gather feedback and iterate on guardrails before the model reaches general product surfaces.
The Cybersecurity Benchmarks, In Detail
Google’s headline numbers cover the full defense lifecycle. On SEC-bench Pro, which measures finding real vulnerabilities in real-world software, Carbon scores 88.2 percent. On ExploitGym, which tests whether a found flaw can actually be exploited, it scores 54.8 percent. On CWE-bench, which measures turning validated flaws into working patches, it reaches 71.0 percent. Google presents these as state-of-the-art results, evaluated without guardrails.
The internal benchmarks point the same way. Carbon generates secure code at more than twice the success rate of Argon on Google’s internal tests, and Google’s Chrome team reports that Carbon performed more than ten times better than another lab’s large frontier model when generating high-quality patches. Wiz, Google’s cloud security platform, also reports Carbon leading on benchmarks for web penetration testing. These are company-reported figures, and they should be read that way until independent evaluations appear.
Beyond security, Carbon is built for sustained agentic work. Google says its internal teams have seen it run for days, persisting through dead ends to root-cause and patch subtle bugs across large codebases. It also jumps 89 Elo points to fourth place, at 1768, on the WebDev Arena benchmark for full-stack web development.
Safety Claims That Invite Scrutiny
Google pairs the capability claims with an unusually detailed safety pitch. Carbon, it says, reduced successful cheating on Terminal-Bench 4.0 to zero percent, with cheating attempts dropping to 0.4 percent, while improving overall accuracy. In live coding environments with Google’s own teams, high-severity misalignment alerts fell by almost 80 percent. Internal testing also found Carbon generating roughly half as many unsupported claims as Argon, putting methodological caveats directly into its deliverables rather than fabricating evidence.
Collaboration under ambiguity is another focus. Rather than guessing when a request is unclear, Carbon retrieves relevant sources and surfaces discrepancies before acting. Google reports a 26 percent reduction in overreach and a 29 percent reduction in under-service, meaning the model both stays within scope more reliably and completes multi-step plans more fully. On prompt injection, Carbon achieves a zero percent attack success rate on Gray Swan’s Indirect Prompt Injection benchmark, improving on Argon’s already low 0.7 percent.
What It Costs, And Who Gets It First
When Carbon does reach developers, it will launch at the same introductory price as Argon: 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens priced at 95 percent off the input price. After the introductory period, the price rises to 4 dollars per million input tokens and 20 dollars per million output tokens. Google has not said when general availability begins.
For now, the practical advice is simple. Security teams interested in early access should ask Google about the Fairwind Program, while everyone else can keep working with Argon, which remains the generally available Gemini 4 model. The two-week gap between Argon and Carbon also raises a planning question for developers: building on a frontier model that may be superseded within a fortnight is a different calculation from adopting a model expected to lead for months, as Gemini Enterprise is already reaching into workplace workflows and DeepMind’s world models show how fast the Gemini family is branching.
Why It Matters
Carbon matters because it turns the frontier-model race toward a specific job: defending software rather than merely writing it. If its benchmark leads hold up, security teams gain a tool that finds, proves and patches flaws in one workflow. But the restricted rollout is the other half of the story: a model this capable, kept this close, concentrates early access in a chosen few defenders, and the rest of the industry must wait for evidence justifying trust in it.
Background
Gemini 4 is Google’s current frontier generation, with Argon opening the line two weeks before Carbon extended it toward cybersecurity and long-horizon engineering. Google has framed both models as part of a faster research cadence in which its own models help advance the next generation, accelerating releases and, with them, the questions about who gets access first.
Source: Google DeepMind announcement
0 Comments