# Gemini 4 Carbon Tops Artificial Analysis's Index, but Google Is Again Giving It to Cyber Defenders First

Times of AI Desk · 2026-10-29 · Models

[https://timesof.ai/2026/10/google-gemini-4-carbon-fairwind-cyber-frontier](https://timesof.ai/2026/10/google-gemini-4-carbon-fairwind-cyber-frontier)

> Google DeepMind announced Gemini 4 Carbon on October 28 and is releasing it first to vetted cyber defenders through Fairwind and to U.S. government pre-release testing, with wider access in phases. Artificial Analysis scored it 59 on its Intelligence Index v5.0, the highest it has measured, but ranked it second on its coding-agent index and measured a higher hallucination rate than Argon. Google's cyber benchmark claims were run without guardrails and are not independently checked.

Google's newest frontier model is the first to sit on top of Artificial Analysis's new Intelligence Index, and almost nobody can use it yet. Gemini 4 Carbon, announced by Google DeepMind's Koray Kavukcuoglu on October 28, goes first to a subset of cyber defenders through Google's Fairwind program and to the U.S. government's voluntary pre-release testing. Developers, businesses and consumers get it in phases afterwards, on dates Google has not published.

That is the same gated route Google used for [Gemini 4 Argon](https://timesof.ai/2026/09/google-gemini-4-argon) and, before it, for the cyber tier of [Gemini 3.8 Flash](https://timesof.ai/2026/09/google-gemini-3-8-flash-cyber-fairwind). The difference this time is that an outside benchmarker has already scored the model, so some of Google's claims can be checked against something other than Google.

## What the independent check found

Artificial Analysis said Carbon scored 59 on its Intelligence Index v5.0, the highest of any model it has tested, and placed first in four of the index's nine evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. Google cites the same number-one ranking in its announcement.

The lead is not uniform. On Artificial Analysis's Coding Agent Index, Carbon scored 59, behind Claude Code running Anthropic's Claude Opus 5.5 at 62. On AA-Briefcase, an agentic knowledge-work evaluation, Carbon's 1609 Elo trailed xAI's Grok 4.7, Meta's Muse Wonder and Moonshot's Kimi K3.1 Pro.

Two other findings cut against the launch framing. Artificial Analysis measured a hallucination rate of 11% for Carbon, up from 4% for Argon, even though Google says Carbon makes about half as many unsupported claims as Argon in its own testing. And Carbon used about 123,000 output tokens per index task; at standard pricing, Artificial Analysis put the cost of running its index at $7.76 per task, against $10.04 for Claude Opus 5.5, $6.25 for OpenAI's GPT-6 Astra and $7.68 for Argon.

## What only Google has measured

The cyber numbers, which are the reason for the gated release, are Google's. Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on Collinear's CWE-bench, and says those cyber evaluations were run without guardrails. It says Chrome's security team found Carbon's patches "more than 10X better" than those of an unnamed other lab's large frontier model, and that Carbon delivers more than twice Argon's uplift on secure code in internal evaluations.

Google also reports a gain of 89 Elo points on WebDev Arena, to fourth place, and safety figures from its own and partner tests: 0.0% cheating on Terminal-Bench 4.0, an 80% drop in high-severity misalignment alerts, and a 0.0% indirect prompt-injection success rate on Gray Swan's benchmark, against 0.7% for Argon.

## Price and access

When Carbon reaches the API, Google lists an introductory price of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20, with cached input 95% cheaper. Ars Technica reported that Google declined to say when the introductory price ends; Artificial Analysis also said there is no confirmed end date. Ars noted that the gaps between Carbon and its rivals are a few points on most tests and raised the possibility that the model is a post-training upgrade of Argon. Google has not described how Carbon was built.

## Limits

- Carbon is not yet available to the public, so outside reproduction is limited to Artificial Analysis's access. Its index scores are one independent check, not a consensus.
- The cyber, safety and WebDev Arena figures are Google's own reported results. We did not check the WebDev Arena or LMArena text boards for this report.
- Google has not given dates for developer, enterprise or consumer access, or for the end of introductory pricing.
- Whether Carbon is a new base model or a further-trained Argon is unconfirmed.

## Sources

- [Google: Introducing Gemini 4 Carbon (October 28, 2026)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- [Artificial Analysis: Gemini 4 Carbon is the new #1 in the Artificial Analysis Intelligence Index (October 28, 2026)](https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1)
- [Ars Technica: Google announces Gemini 4 Carbon, its latest frontier AI model that you can't use (October 28, 2026)](https://arstechnica.com/ai/2026/10/google-announces-gemini-4-carbon-its-latest-frontier-ai-model-that-you-cant-use/)