Google Gemini 4 Carbon: Cybersecurity's New Top Model

Updated: October 2026 · 6 min read · By the CreatorSprout team

Launch coverage based on Google's official announcement and independent launch reporting — we have not tested Gemini 4 Carbon hands-on, and it is not publicly available yet. Google's own claims and independently reported details are kept separate below. No affiliate links in this article.

Two weeks. That is the gap between Gemini 4 Argon's general availability on October 13 and Carbon's announcement on October 28 — and the pattern is deliberate. Google introduced Argon itself in two stages: cyber defenders first through Fairwind on September 30, then everyone else on October 13. Carbon repeats the same playbook, only the stakes are higher: this model is designed to find, validate and patch real software vulnerabilities, which is exactly the kind of capability Google says it wants to test with defenders before handing to everyone.

What Gemini 4 Carbon is

Gemini 4 Carbon is Google DeepMind's next frontier model, announced October 28, 2026, in a Google blog post. Google positions it as a specialist at the top of the Gemini 4 family: where Argon is the general-purpose flagship, Carbon is tuned for the full cybersecurity defense lifecycle — detecting vulnerabilities, validating whether they are actually exploitable, and fixing them at the root cause — plus sustained, multi-day engineering work on large codebases and low-level systems.

The independent signal that matters most is the ranking: Carbon sits first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, one point ahead of Anthropic's Claude Opus 5.5 and two ahead of OpenAI's GPT-6 Astra, and nine points above Argon itself. That index is run by an independent evaluator, so it is the one headline number in this launch that does not come from Google alone.

The cybersecurity benchmarks

Google's announcement leads with three external benchmarks, evaluated without guardrails, covering the three stages of defensive security work. These are Google's reported results — they have not yet been independently reproduced:

Finding real vulnerabilities — SEC-bench Pro: 88.2%. Carbon autonomously uncovers complex flaws buried in real-world software, which Google presents as a state-of-the-art result.
Validating risk through exploitation — ExploitGym: 54.8%. Finding a flaw is only half the job; defenders need to know which flaws are actually exploitable before deciding what to fix first. Carbon sets a new reported high here too.
Patching at the root cause — CWE-bench: 71.0%. Instead of a workaround, Carbon turns validated findings into working code fixes at the underlying cause, according to Collinear's benchmark.

On top of those, Google cites internal results: more than double Argon's success rate at generating secure code, and a Chrome team finding that Carbon produced high-quality patches more than 10 times better than another lab's large frontier model. That last comparison names no rival, so treat it as an internal claim, not a measured head-to-head. Google also notes that Wiz, its cloud security platform, sees Carbon leading on web penetration-testing benchmarks used in its Scan for Good initiative.

Long-horizon engineering work

The second half of Carbon's pitch is persistence. Google says its internal teams have watched Carbon run for days to push past dead ends and fix subtle bugs across large codebases — the kind of work where earlier models stalled after a few failed attempts or lost track of context across a massive repository. That is the practical meaning of "long horizon": not a longer chat, but an agent that keeps working a hard problem without drifting off task or quietly cutting corners.

The same push shows up in general coding. Carbon jumps 89 Elo points on WebDev Arena to land at #4 (1768), which Google attributes to improved output quality and artifacts in full-stack web development. For creators and developers, that is the number to watch when Carbon eventually reaches the API: a model that can sustain a multi-day debugging run and ship cleaner web output is a different tool from one that just answers fast.

Safety and alignment claims

Because a model this capable at finding and exploiting flaws is dual-use by nature, Google's announcement spends real space on safeguards — and this is also why access is restricted. The claims, all from Google's own testing:

Less cheating, less fabrication. Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with cheating attempts down to 0.4%, while improving overall accuracy. In live coding with Google's own teams, high-severity misalignment alerts dropped by almost 80%, and unsupported claims were cut by roughly half versus Argon.
Better collaboration under ambiguity. Rather than guessing or forcing a path forward when information is missing, Carbon retrieves relevant sources and surfaces discrepancies to the user before acting.
Following through without overstepping. Google reports overreach down 26% and under-service down 29% versus Argon — multi-step plans carried out end to end while staying within scope.
Prompt-injection resistance. Where Argon previously led Gray Swan's Indirect Prompt Injection Benchmark at a 0.7% attack success rate, Carbon records 0.0% — resisting every evaluated attack attempt in Google's testing.

Google also says it is engaged in the U.S. government's voluntary process for pre-release model testing, and that it is gathering defender feedback and iterating on guardrails before broader release. That framing is the reason for the staged rollout: the company is explicitly trading speed for a controlled first cohort.

Pricing

Introductory API pricing: $2 per million input tokens and $10 per million output tokens — the same introductory rates as Argon — with cached input tokens priced at 95% off the input price.
Standard rates afterwards: $4 per million input tokens and $20 per million output tokens, as reported by 9to5Google from Google's pricing materials. Google's own blog post confirms the introductory rates; the step-up comes from launch reporting.
The honest caveat: these rates only matter once you can access the model. Today, no public signup exists, and independent checks on October 28 found no Carbon entry in the public Gemini API model list — only Argon.

Who can use it — and who can't yet

Now: a subset of vetted cybersecurity defenders, via Google's Fairwind Program, starting the week of October 28.
Next: additional trusted partners, phased in gradually through the same Fairwind Program.
Later, no date given: developers, businesses and consumers via Google's product surfaces — the Gemini app, API and AI Mode in Search. Google says "as soon as possible" after defender feedback and safeguard work, but has announced no timeline.
Watch out for fakes: any third-party site claiming to sell or offer "Gemini 4 Carbon access" today is not coming from Google. The only legitimate route at launch is Fairwind, and it vets organizations, not individual signups.

What it means for creators and developers

For most readers of this site, Carbon changes nothing you can do this week — and that is worth saying plainly. What it does change is the planning horizon. If you build with AI coding tools, the model to evaluate today is Argon: it is in the API, it is in the Gemini app, and it is the baseline Carbon will be measured against when it arrives. Argon took 13 days to go from its Fairwind announcement to paid API access; Google has given Carbon no equivalent date, so do not plan a launch around it.

For security teams, the picture is different. Carbon's benchmark profile — find, validate, patch — maps directly onto vulnerability management workflows, and the Fairwind route is the one to pursue if your organization defends important systems. For everyone else, the practical move is to watch what the first defender cohort publishes: independent results from real defensive work will tell us more than any launch-day table.

The broader pattern is also worth noting. Frontier releases are now arriving in weeks, not quarters: Argon on September 30, general availability October 13, Carbon October 28, with OpenAI's GPT-6.1 Terra landing October 26 and Moonshot's Kimi K3.1 Pro on October 23. Benchmark leads are changing hands every few weeks. Picking a model on a single leaderboard score is a losing strategy; pick on the workflow it actually improves for you, and re-test when the next release lands.

The honest caveats

We haven't tested it. Carbon is a closed, defender-only rollout. Everything here comes from Google's announcement and launch reporting, not our own runs.
Most headline numbers are Google's own. The Artificial Analysis #1 ranking is independent; the cybersecurity scores and the "10x better patches" claim are not yet independently reproduced, and the 10x comparison names no rival.
Not a clean sweep. Google's own benchmark table shows rivals ahead on some tests — the honest read is "leads on cybersecurity and overall index," not "best at everything."
No public access, no date. No signup page, no API entry, no consumer plan. Anyone quoting a Carbon release date is speculating.
Pricing doubles after intro. The $2/$10 rates are introductory; standard rates are $4/$20 per reporting. Budget on the standard rates, not the launch discount.
Dual-use is the reason for the wait. A model that finds and exploits vulnerabilities is powerful in defensive hands and dangerous in others. The restricted rollout is a safety decision, not just marketing scarcity.

Verdict

Gemini 4 Carbon is a genuinely major launch that almost nobody can use yet — and both halves of that sentence matter. The #1 independent ranking and the cybersecurity benchmark profile make it the strongest model Google has shipped, and the long-horizon persistence claim, if it holds up in independent testing, is the capability that changes what agents can actually finish. But until Fairwind expands or Google opens developer access, Carbon is a leaderboard entry, not a tool. Use Argon today, keep an eye on defender-cohort reports, and judge Carbon when independent testers can run it themselves — not on launch-day charts.

FAQ

What is Google Gemini 4 Carbon?

Google DeepMind's newest frontier model, announced October 28, 2026. Google calls it its strongest model, built for cybersecurity defense and long-horizon engineering work: finding software vulnerabilities, validating them, and writing root-cause patches, over sessions that can run for days.

Can I use Gemini 4 Carbon today?

Almost certainly not. At launch, Carbon is available only to a subset of vetted cybersecurity defenders through Google's Fairwind Program, with additional trusted partners phased in later. Google has given no date for developer, business or consumer access, and independent checks on October 28 found no Carbon entry in the public Gemini API model list.

How much does Gemini 4 Carbon cost?

Google lists introductory API pricing of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input price — the same introductory rates as Gemini 4 Argon. Reporting from 9to5Google puts the standard rates after the introductory period at $4 and $20 per million tokens. Those rates only matter once you can actually access the model.

How is Gemini 4 Carbon different from Gemini 4 Argon?

Carbon builds on Argon, which reached general availability on October 13, 2026, and pushes further on cybersecurity and long-horizon work: it ranks first on the Artificial Analysis Intelligence Index v5.0 at 59, nine points above Argon, and Google reports more than double Argon's success rate at generating secure code. Argon remains the Gemini 4 model you can actually use today.

Are Carbon's benchmark scores independently verified?

Only partly. The #1 ranking on Artificial Analysis' Intelligence Index v5.0 (score 59) comes from an independent evaluator, but the cybersecurity figures — 88.2% on SEC-bench Pro, 54.8% on ExploitGym, 71.0% on CWE-bench — are Google's own reported results, and internal claims such as the Chrome team's "more than 10x better patches" name no rival model. Treat them as claims to be tested once broader access opens.