What Is Gemini 4 Carbon?
On October 28, 2026, Google DeepMind announced Gemini 4 Carbon, calling it the company’s strongest AI model to date. Carbon is a frontier model designed for two things: finding and fixing software security flaws, and handling long, complicated engineering tasks that take days rather than minutes.
Carbon arrives just two weeks after Gemini 4 Argon, and Google says it builds directly on Argon’s foundation. The difference is focus. Where Argon was aimed at general hard engineering problems, Carbon is tuned specifically for the cybersecurity defense lifecycle: detecting vulnerabilities, proving they are real, and patching them at the root cause.
The announcement came from Koray Kavukcuoglu, SVP of Google DeepMind and Google’s chief AI architect, in an official Google blog post.
Gemini 4 Carbon Benchmark Scores
Google published a set of benchmark results with the launch. These are Google’s own numbers, so they should be read with that in mind — independent verification is still pending.
| Benchmark | What it tests | Carbon’s score |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | Overall model intelligence | 59 — ranked #1 |
| SEC-bench Pro | Finding real vulnerabilities in software | 88.2% |
| ExploitGym | Turning a disclosed flaw into a working exploit | 54.8% |
| CWE-bench | Patching bugs at the root cause | 71.0% |
| WebDev Arena | Web development quality | #4 (1768 Elo, +89 points) |
Google also reported internal results: Carbon more than doubles Argon’s success rate at generating secure code, and Google’s Chrome team says Carbon produced high-quality patches more than 10 times better than another lab’s large frontier model. Wiz, Google’s cloud security platform, reports Carbon leading in web penetration testing for its “Scan for Good” initiative.
Built for Long-Horizon Engineering Work
One of the most interesting claims about Carbon is about endurance. Google says its internal teams have run the model for days at a time on multi-step engineering tasks — root-causing subtle bugs across large codebases and low-level systems — without it losing the thread. This is what researchers call “long-horizon” capability: the model can persist through dead ends instead of giving up after a few steps.
On safety, Google says Carbon improves on Argon’s already strong record. Successful cheating attempts on the Terminal-Bench 4.0 benchmark dropped to 0.0%, high-severity misalignment alerts in live coding fell by almost 80%, and the model resisted every evaluated attack in the Gray Swan indirect prompt injection test (a 0.0% attack success rate, down from 0.7%).
Why Cyber Defenders Get It First
The most unusual part of this launch is not the model — it is the release order. Starting this week, Carbon is going first to a subset of cyber defenders: a small group of vetted cybersecurity organizations. Access then expands in phases to additional trusted partners through Google’s Fairwind Program, which already has more than 650 partners, with priority for governments, national cyber authorities, and critical infrastructure operators.
Google is also taking part in the U.S. government’s voluntary process for pre-release model testing. Only after gathering feedback and improving safeguards will Carbon reach developers, businesses, and consumers through Google’s products “as soon as possible.” No general availability date has been given.
The logic is straightforward: a model this good at finding software flaws could be misused, so Google wants defenders to have it before attackers do. It is a notable shift from the usual race to put every new model in everyone’s hands immediately.
How Carbon Compares to Other Frontier Models
Google’s headline claim is that Carbon ranks #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59, up from Argon’s 50 and just ahead of rivals like Claude Opus 5.5 and GPT-6 Astra. On security benchmarks, Google’s charts place OpenAI’s GPT-6 Astra second on SEC-bench Pro at 85.4%, with Argon at 83.4% and Carbon at 88.2%.
| Model | AI Index v5.0 score | SEC-bench Pro | Availability |
|---|---|---|---|
| Gemini 4 Carbon | 59 (#1) | 88.2% | Limited: cyber defenders only |
| Gemini 4 Argon | 50 | 83.4% | Trusted partners |
| GPT-6 Astra (OpenAI) | Below 59 | 85.4% | Public |
| Claude Opus 5.5 (Anthropic) | Below 59 | Not disclosed | Public |
Remember: these are Google’s own reported numbers. Independent evaluations from Artificial Analysis and others will confirm or challenge them in the coming weeks, and we will update this post when they do.
Gemini 4 Carbon Pricing
Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced 95% below the input price. After the introductory period, standard pricing rises to $4 per million input tokens and $20 per million output tokens.
Note that reports differ on exactly how long the introductory price lasts, because Google has not given a public end date. Treat the lower prices as temporary.
When Can You Use Gemini 4 Carbon?
Honestly: not yet, unless you work at a vetted cyber defense organization. There is no public API model ID for Carbon, no published context window or model card, and no confirmed date for wider access. If you are a developer waiting to build with it, the most realistic expectation is weeks to months, depending on how the early phases go.
What you can do now: keep an eye on Google’s developer channels and the Fairwind Program page for expansion announcements. If you work in security, the Fairwind Program is the route to request early access.
What This Means for You
For most readers, Gemini 4 Carbon changes nothing today — you cannot download it or try it. What it signals is bigger: frontier models are now being built specifically for defense, and labs are starting to release the most powerful ones to trusted users first rather than everyone at once. That is a meaningful shift in how AI reaches the public.
For developers and security professionals, Carbon’s benchmark numbers — if they hold up under independent testing — suggest the next generation of coding assistants will be dramatically better at writing secure code and reviewing it. For everyone else, the practical takeaway is that the AI tools arriving in 2027 will be built on models like this one.
The Bigger Picture: Defense-First AI Is a New Trend
Carbon is not arriving in a vacuum. The same week Google announced it, Mistral released the weights of its 1-trillion-parameter Large 4 model as open weights, and CrowdStrike reported attackers using AI agent frameworks against more than 200 organizations. The pattern is clear: AI that finds and fixes software flaws is now one of the most strategically important capabilities in the industry.
What makes Carbon’s launch different is the deliberate slowness. Most frontier releases race to reach every developer on day one. Google is doing the opposite — giving the model to defenders first, then partners, then everyone else. As Ars Technica summarized the launch: the most capable new model of the month is one “you can’t use” yet.
That frustration is real, but the reasoning is worth taking seriously. A model that scores 88.2% at finding real vulnerabilities and 54.8% at turning them into working exploits is genuinely dual-use: the same capability that lets a security team patch a flaw lets an attacker weaponize it. Google’s bet is that defenders should get the head start this time.
Whether other labs follow this playbook will be one of the most interesting questions of 2027. If Anthropic and OpenAI start staging their most powerful releases through trusted programs too, the era of instant public access to every new model may be ending.
Summary
Google’s Gemini 4 Carbon, announced October 28, 2026, is the company’s strongest AI model yet — leading key benchmarks in vulnerability detection, exploitation, and patching, and ranked #1 on the Artificial Analysis Intelligence Index. It is rolling out first to cyber defenders through the Fairwind Program, with wider access to follow. Pricing starts at $2/$10 per million tokens during the introductory period. The launch marks a new era of defense-first AI deployment.
For background on how governments are shaping AI policy around releases like this, see our guides to America’s AI Action Plan and Japan’s AI regulation framework.
Sources
- Google Blog — Introducing Gemini 4 Carbon (official announcement)
- WebProNews — Google Gives Cyber Defenders First Access to Gemini 4 Carbon
- Aragon Research — Gemini 4 Carbon Wins Benchmarks, But Buyers Can’t Use It Yet
What is Gemini 4 Carbon?
Gemini 4 Carbon is Google DeepMind’s newest frontier AI model, announced on October 28, 2026. It is designed for cybersecurity defense and long-horizon engineering tasks, and Google calls it the company’s strongest model to date.
When will Gemini 4 Carbon be available to the public?
There is no confirmed date yet. Google is first giving Carbon to a subset of cyber defenders, then expanding to trusted partners through its Fairwind Program. Developers, businesses, and consumers will get access as soon as possible after safeguards are improved.
How much does Gemini 4 Carbon cost?
Carbon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached inputs 95 percent cheaper. Standard pricing rises to $4 per million input and $20 per million output after the introductory period, though Google has not announced when it ends.
What benchmarks does Gemini 4 Carbon lead?
According to Google, Carbon ranks number 1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59, and leads three security benchmarks: 88.2 percent on SEC-bench Pro, 54.8 percent on ExploitGym, and 71.0 percent on CWE-bench. These are Google’s own reported results and await independent verification.
Is Gemini 4 Carbon better than GPT-6 Astra or Claude Opus 5.5?
On Google’s reported numbers, Carbon leads the overall AI Index and the security benchmarks. However, these figures come from Google itself and have not yet been independently verified, so independent tests in the coming weeks will give the final answer.
Ahesanur Asun is the Founder and Editor of UptoFeed, where he publishes updates on
AI news, artificial intelligence tools, and emerging technology. He focuses on explaining complex AI developments in a clear and practical way so readers can quickly understand and use the latest innovations.
