# Gemini 4 Carbon vs Argon: which model security teams should pick

> Google's Gemini 4 Carbon beats Argon on cyber benchmarks and alignment, but access is limited to a few defenders for now.

- Published: 2026-10-29
- Topic: [AI](https://futuretechnologyhq.com/topic/ai/)
- URL: https://futuretechnologyhq.com/article/gemini-4-carbon-vs-argon-which-model-security-teams-should/
- Publisher: [Future Technology HQ](https://futuretechnologyhq.com/), free to read. Quote as "Future Technology HQ" with a link to the article.

## Key points

- Carbon scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, and Google claims more than 2x Argon's success rate for generating secure code
- Access is limited to a subset of cyber defenders through the Fairwind Program, so most teams cannot yet choose between the two
- The $2 and $10 per million token introductory price rises to $4 and $20 after the introductory period, so budgets should assume the higher rate

## Article

Security teams weighing Google's latest models face a choice that is partly academic: Gemini 4 Carbon, announced on 28 October 2026, beats its two-week-old sibling Gemini 4 Argon on nearly every cyber measure Google published, but only a small group of defenders can use it. Argon is already available to developers, while Carbon is limited to a subset of cyber defenders for now.

## The contenders

Gemini 4 Argon launched two weeks before Carbon and Carbon builds on its safety and security stack. Carbon continues that approach. Google says it is rolling Carbon out to a subset of cyber defenders this week and will phase in more trusted partners through the same Fairwind Program, while it gathers feedback and tunes guardrails before opening it to developers, businesses and consumers.

Carbon is built for long-horizon work, meaning tasks that run for days rather than a single prompt. Google says internal teams have watched it persist through dead ends to root-cause and patch subtle bugs across large codebases and low-level systems. It also ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59.

## Head-to-head comparison

Google has published direct comparisons for some measures and absolute figures for others. Where the source gives no Argon number, the table says so rather than guessing.

<div class="comparison-table">
<table>
<thead><tr><th>Feature</th><th>Gemini 4 Carbon</th><th>Gemini 4 Argon</th></tr></thead>
<tbody>
<tr><td>Availability</td><td>Subset of cyber defenders, phasing to trusted partners via Fairwind</td><td>Available to developers</td></tr>
<tr><td>Introductory price</td><td>$2 per million input tokens, $10 per million output tokens</td><td>Same introductory price</td></tr>
<tr><td>Price after introductory period</td><td>$4 input, $20 output per million tokens</td><td>Not stated in the Carbon announcement</td></tr>
<tr><td>Secure code generation (internal benchmark)</td><td>More than 2x Argon's success rate</td><td>Baseline</td></tr>
<tr><td>Unsupported claims</td><td>About 2x fewer than Argon</td><td>Baseline</td></tr>
<tr><td>Indirect prompt injection attack success (Gray Swan)</td><td>0.0%</td><td>0.7%</td></tr>
<tr><td>Vulnerability finding (SEC-bench Pro)</td><td>88.2%</td><td>No figure given</td></tr>
<tr><td>Exploitation (ExploitGym)</td><td>54.8%</td><td>No figure given</td></tr>
<tr><td>Patching (CWE-bench)</td><td>71.0%</td><td>No figure given</td></tr>
</tbody>
</table>
</div>

One caveat applies to the benchmark rows: Google says the external cyber scores were evaluated without guardrails, so they show raw capability rather than what a deployed product will do.

## Where Carbon wins

The headline is cyber defence across the full lifecycle: finding, validating and fixing flaws. The 88.2% on SEC-bench Pro covers finding real vulnerabilities, 54.8% on ExploitGym covers checking whether they are exploitable, which helps teams decide what to fix first, and 71.0% on Collinear's CWE-bench covers turning findings into working patches.

Google also cites outside validation, though these are partner statements rather than independent audits. The Chrome team reported that Carbon produced high-quality patches more than 10 times better than another lab's large frontier model. Google says Wiz, its cloud security platform, observes Carbon leading on web penetration testing benchmarks, a core capability for Wiz's Scan for Good initiative.

The behavioural gains matter as much for anyone letting an agent loose on production code. Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts at 0.4%, while improving accuracy. High-severity misalignment alerts in live internal coding environments fell by almost 80%. Overreach dropped 26% and under-service 29%, which is a dry way of saying the model is more likely to do what it was asked, no more and no less. It also scored 0.0% on Gray Swan's prompt injection benchmark, where an attacker hides instructions in content the model reads, against Argon's 0.7%. For web development, Carbon rose 89 Elo points to fourth place on WebDev Arena at 1768.

## Where Argon wins

Argon's advantage is simple: it exists in more hands. Carbon's wider release has no date, and Google describes it only as coming "as soon as possible". A team that needs a capable model for a project this quarter cannot plan around Carbon.

Argon also has a published Safety and Security report. Google published its propensity to behave in an aligned way in a Safety and Security report, and Carbon is described as building on the same safety stack of model training, real-time misuse defences, prompt-injection resilience and chain-of-thought monitoring. Argon's pricing mechanics are covered in our piece on [Argon's $2 per million token price](/article/gemini-4-argon-pricing-2-dollars-per-million-tokens/).

## UK and US implications

For US organisations, Google says it is taking part in the US government's voluntary pre-release testing process, which may matter to firms with procurement or disclosure obligations. UK security teams face the same access question without a stated route: Google has not announced UK availability or pounds pricing for Carbon; the introductory price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20.

## The verdict

Security teams already inside Fairwind should test Carbon first, particularly for vulnerability triage and patch generation, where Google's figures are strongest. Everyone else should keep building on Argon and treat Carbon as the upgrade to plan for, not a purchase to make.

Budget for the post-introductory $4 and $20 rates rather than the $2 and $10 launch price; the cost-per-task maths in our [comparison of frontier model bills](/article/sonnet-5-5-gpt-6-1-sol-gemini-4-argon-cost-per-task/) shows why headline token prices mislead. Cached input tokens at 95% off will matter for agents that re-read large codebases. Treat the vendor-reported figures as claims to verify in a pilot, not procurement evidence.

## Key takeaways

Carbon is Google's stronger model on cyber defence, alignment and prompt-injection resistance, by Google's own measurements. Carbon is available only to a subset of cyber defenders for now. The practical step is to apply for access if the work is security-critical, and to run a controlled pilot against internal code before trusting any benchmark number.

## Sources

- [Google Blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- [9to5google.com](https://9to5google.com/2026/10/28/gemini-4-carbon-announcement/)
- [constellationr.com](https://www.constellationr.com/insights/news/google-rolls-out-gemini-4-carbon)
