# Google's Gemini 4 Carbon tops an intelligence index and posts strong security scores

> Google says Gemini 4 Carbon scored 59 on Artificial Analysis Intelligence Index 5.0, ahead of Claude Opus 5.5 and GPT-6 Astra. Access starts with a few cyber defense firms, at $2 and $10 per million tokens.
>
> OpenCheese AI · 2026.10.29 · Web page: https://opencheese.dev/blog/gemini-4-carbon-tops-intelligence-index-with-strong-security-scores

Google says its new model, Gemini 4 Carbon, now sits first on the Artificial Analysis Intelligence Index 5.0, with 59 points. The company then hands the model, at first, to a short list of cyber defense firms rather than to everyone who would like to try it.

Carbon is aimed at programming, long multi-step tasks and cybersecurity. Everything below comes from Google's own claims, and the report quotes no outside verification.

## The ranking claim

According to Google, Carbon's 59 points put it ahead of two named rivals, Claude Opus 5.5 from Anthropic and GPT-6 Astra from OpenAI. The report gives no scores for those two models, so the size of the gap is unknown. First place by one point and first place by ten look identical in a press summary.

One detail is worth keeping straight. The ranking is a general intelligence index. The report does not say the rivals were beaten on the security tests specifically, only that Carbon topped the overall list.

## The security numbers

Google published three security benchmark results for Carbon:

| Benchmark | What it measures | Carbon's result |
|---|---|---|
| SEC-bench Pro | Finding vulnerabilities in real software | 88.2% |
| ExploitGym | Exploiting the bugs that were found | 54.8% |
| CWE-bench | Fixing vulnerabilities | 71% |

These are three different tests with different difficulty, so reading them as a single funnel would be generous. Still, the pattern is plausible. Spotting a flaw is the easy part, turning it into a working exploit is harder, and the report gives Carbon a lower score on that step than on discovery.

A model that can find, exploit and patch bugs is useful to defenders and equally interesting to everyone else. The report does not say whether that is why the rollout is restricted, but the guest list suggests Google noticed.

## Gains over Gemini 4 Argon

Google compares Carbon with its predecessor, Gemini 4 Argon, and claims several improvements:

- **Secure code.** The success rate of generating secure code more than doubled. No baseline figure is given, so "more than double" has no number attached to either side.
- **Long engineering work.** Carbon handles extended engineering tasks better.
- **Fewer unsupported claims.** The model makes fewer statements it cannot back up, which is a polite way to describe a familiar habit of language models.
- **Prompt injection.** It resists attacks that hide malicious instructions inside the requests it receives.

The last item matters for any model that reads untrusted text while doing a job. Google gives no test or score for it here, only the claim.

## Access and pricing

In the first phase, Carbon goes only to select companies that specialize in cyber defense. Google then plans to widen the circle of partners through a program called Fairwind. The report gives no dates and no criteria for who gets in next.

Pricing at launch is $2 per million input tokens and $10 per million output tokens. For a model positioned around long engineering tasks, where output can run long, the second number will probably be the one that shows up on invoices. The report does not say whether those prices apply to the initial partners or only to later, broader access.

## What the report leaves out

Several things are missing, and they are the things a skeptical reader would ask for first:

- Scores for Claude Opus 5.5 and GPT-6 Astra on the same security benchmarks.
- Any independent replication of Google's figures.
- A timeline for general availability.
- Details of how Fairwind works or who can join.
- Baselines behind the "more than doubled" claim.

Until those appear, Carbon is a model with strong stated numbers and a controlled debut. For now, the guest list is shorter than the benchmark table.

## Related posts

- 2026.10.29 · [Google's Gemini 4 Carbon tops an intelligence index but ships first to cyber defenders](https://opencheese.dev/blog/google-gemini-4-carbon-tops-intelligence-index-cyber-defenders-first.md): Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a 59 and posts leading cybersecurity scores. For now, only a subset of cyber defenders can use it.
- 2026.10.14 · [Google launches Gemini 4 Argon, priced at about a fifth of rival frontier models](https://opencheese.dev/blog/google-launches-gemini-4-argon-priced-at-a-fifth-of-rival-frontier-models.md): Google DeepMind has released Gemini 4 Argon, claiming it beats GPT-6 Astra and Claude models on coding, security, science, finance and law. Introductory pricing is $2 and $10 per million tokens.
- 2026.10.29 · [Chrome's enhanced autofill can now pull order numbers and bookings from Gmail](https://opencheese.dev/blog/chrome-s-enhanced-autofill-can-now-pull-order-numbers-and-bookings-from-gmail.md): Google says Chrome's enhanced autofill will use Gemini with Personal Intelligence to suggest form entries from Gmail, such as order numbers and flight bookings. It starts in the US, with settings to limit or disable it.