For two weeks, the strongest AI model Google has ever built was deliberately kept away from almost everyone — handed only to a small circle of cybersecurity defenders, while the rest of the world waited.
That wait ended today. Gemini 4 Carbon is now generally available, and in this full breakdown we will explain what it is, what the benchmark numbers actually mean, what it costs, who can use it — and the one honest catch hiding inside its record-breaking safety score, which we will save for the final third.
What Is Gemini 4 Carbon?
Gemini 4 Carbon is Google DeepMind's next frontier AI model, announced on October 28, 2026, as the successor to Gemini 4 Argon, which had launched only two weeks earlier.
Google positions the two models side by side. Argon remains the recommended choice for the majority of software engineering and knowledge work. Carbon is built for the tasks that test the limits of a model's intelligence: complex multi-step coding and debugging, long-horizon engineering work that can run for extended stretches, and open-ended scientific research. To do that, Carbon deliberately uses more tokens than Argon — exploring more thoroughly, reasoning more deeply, and running extended verification loops before it answers.
It is a reasoning model with a 1 million token context window. It accepts text, image, audio and video input and produces text output.
Why Cyber Defenders Got It First
The strangest part of this launch is the release order, and it tells you what Carbon is really for.
Instead of a normal public release, Google first gave Carbon to a subset of cybersecurity defenders through its Fairwind Program, and engaged in the U.S. government's voluntary process for pre-release model testing. Only after two weeks of defender feedback and further guardrail work did general availability follow.
The reason is simple: Carbon is Google's strongest model for cybersecurity, across the full defensive lifecycle — finding vulnerabilities, proving whether they can actually be exploited, and patching them at the root cause. A model that good at offence-adjacent skills has to be placed in defenders' hands first.
Gemini 4 Carbon Benchmarks: The Numbers That Matter
Benchmarks are easy to inflate, so let us stick to the headline results Google published and the independent index that ranks all major models together.
Artificial Analysis Intelligence Index v5.0: 59 — rank #1 among all models, ahead of Gemini 4 Argon at 50.
SEC-bench Pro: 88.2% — autonomously discovering complex flaws in real-world software.
ExploitGym: 54.8% — validating whether identified vulnerabilities are genuinely exploitable.
CWE-bench: 71.0% — turning findings into root-cause code fixes.
Terminal-Bench 4.0: 65.2% for agentic terminal and coding work, up from Argon's 57.1%.
Long context: 96.2% F1 on GraphWalks across 256k to 1M tokens, far ahead of Argon's 84.2%.
Google also reports Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than twice lower cost per task than the next best external model — and on the new Multimodal Knowledge Agent benchmark, built with Artificial Analysis for agents working over dense files full of charts and tables, it ranks #2 at more than six times lower cost per task than the top model.
One honesty note: the cybersecurity scores were measured in evaluations run without guardrails, and most headline figures are Google's published results. Independent testing will sharpen the picture over the coming weeks.
Ask yourself this: if a model is this strong at finding flaws in code, would you rather it worked for your security team — or be available to everyone at once? Google's staged rollout was its answer to that question.
Carbon at Work: Real Results From Early Testers
Launch-day claims are one thing. Google's availability announcement includes early field results worth knowing:
In Project Fortify, Google's autonomous system for scanning, triage and remediation, Carbon hit 100% recall on scanning and 95% precision on triage on Google's internal benchmark, and improved open-source code scanning recall by over 25% relative to Argon — surfacing real logic flaws that had gone undetected.
Google's Cloud Vulnerability Research team used Carbon to prove that two bugs in a popular open-source project were in fact exploitable vulnerabilities, serious enough for maintainers to reclassify them as critical and patch quickly.
In Wiz's Scan for Good program, Carbon uncovered zero-day vulnerabilities in two major SaaS providers in under two hours — work Google describes as compressing weeks of vulnerability research into a single afternoon.
On the science side, researchers used Carbon with a new framework called AlphaProof Code to help mathematically prove security properties of the actual code behind the Signal messaging protocol's cryptography — generating millions of lines of proof code in weeks, work that historically takes expert teams years. In protein research, Carbon increased the number of useful evolutionary matches retrieved for hard, low-homolog protein families by 40-fold on average.
The /swarm Feature: One Prompt, a Team of Agents
For developers, the most practical new idea may be /swarm, now in preview in Google Antigravity, Google's agent-first development platform.
Instead of one agent working alone, /swarm turns a single prompt into a coordinated team of agents you can track in real time. You can assign Carbon to the roles that demand the deepest reasoning — architectural design, coordination, auditing — while faster, lighter models work in parallel isolated workspaces, with adversarial verifier agents testing every change.
Combined with Carbon's long-horizon design — built for multi-step engineering tasks such as root-causing and patching bugs across large codebases — this is Google's clearest picture yet of how frontier models will be used: not as chatbots, but as senior members of an agent team.
Safety: The 0.0% Result, and Its Honest Catch
Now the catch we promised at the start — because Carbon's safety story is genuinely impressive, and genuinely incomplete, at the same time.
The impressive part first. On the Gray Swan indirect prompt injection benchmark, Carbon achieved a 0.0% attack success rate — the first time that milestone has been reached, down from 0.7% for Argon. In simulated cyberattack environments, its safeguards blocked 99% of attack trajectories from an adaptive red-teaming agent, cutting the per-attempt breach rate to 0.8%, down from 3.1% on Argon. External testing by Trajectory Labs found zero new universal jailbreaks across 132 hours and more than 13,000 requests. On Google's 375-task agentic honesty benchmark, Carbon passed 92% of tasks, versus 70% for Argon — declining to invent missing experimental data 98% of the time.
Here is the honest catch: in the same announcement, Google itself states that prompt injection robustness "remains an area for continued improvement," and that its own advanced red teaming shows models can still be pushed further against sophisticated attacks in some distributions. In other words, 0.0% on one benchmark is a milestone — it is not a promise that the model cannot be tricked anywhere, ever. Any headline that reads it that way is going further than Google itself does.
Gemini 4 Carbon Pricing and How to Access It
As of November 10, 2026, Gemini 4 Carbon is generally available on the paid tier:
Developers: via the Gemini API in Google AI Studio and Android Studio, and in Google Antigravity for Google AI Ultra subscribers.
Enterprises: via Gemini for business, Google Antigravity, or CodeMender.
Consumers: rolling out in the Gemini app and AI Mode in Search for Google AI Ultra subscribers.
Scientists: Google is expanding its program giving researchers free access to Argon and Carbon at Antigravity's highest usage tier for six months — 10,000 scientists were invited under the original program.
Cyber defenders: continued rolling access through the Fairwind Program.
Pricing starts at the same introductory rate as Argon: $2 per million input tokens and $10 per million output tokens, with cached input priced 95% below the input rate. Per Google's announcement, that introductory pricing is time-limited — the standard rate after the introductory period is $4 per million input and $20 per million output tokens, so heavy users should check the live pricing page before budgeting a big migration.
There is no free tier for Carbon. It is a paid frontier model, and on NEXA AI Tools we will always say that plainly: strongest does not mean free.
Gemini 4 Carbon Explained — the NEXA verdict: this is Google's strongest model by its own published numbers and the independent Intelligence Index, with a defenders-first rollout that was the right call for a cyber model this capable. The 0.0% prompt-injection result is a real milestone, not a magic shield — and the introductory $2/$10 pricing will not last, so the smart move is to test it on your hardest long-horizon work now, while the discount window is open.
Final Word
Gemini 4 Carbon explained in one line: it is the deep-reasoning, cyber-strong, long-horizon member of the Gemini 4 family — the model Google sends in when Argon reaches its limits. You can read Google DeepMind's complete official availability announcement here: https://deepmind.google/blog/gemini-4-carbon-is-now-available/ — and tell us in the comments: if an AI agent could work on one hard problem for you for days without stopping, what problem would you give it first?
Comments 0
Sign in to join the conversation.
No comments yet. Be the first to say something.