Gemini 4 Carbon tops the Artificial Analysis index by one point, but only cyber defenders can use it

Gemini 4 Carbon now leads the Artificial Analysis Intelligence Index, yet it is open only to a subset of vetted security teams. Google DeepMind released Gemini 4 Carbon on October 28, 2026, two weeks after it opened Gemini 4 Argon to the public. The headline is a one-point lead. The sharper story is that most of what Google says about Carbon's cyber skill cannot yet be tested by anyone outside the program.

What Google announced

Google's announcement calls Carbon its strongest model, with "state-of-the-art cybersecurity and long horizon capabilities." The post is credited to Koray Kavukcuoglu, Google's chief AI architect, and it leans on three security numbers. Carbon scores 88.2% on SEC-bench Pro, which measures finding vulnerabilities. It scores 54.8% on ExploitGym, which measures exploiting them. And it scores 71.0% on CWE-bench, which measures patching. Google says the external security evaluations were run without guardrails, so the restricted version people actually get may score differently.

Google also claims more than double Argon's success rate at secure code generation, and says its Chrome team found Carbon more than 10 times better than another lab's large frontier model at producing high-quality patches. Neither figure comes with underlying data in the post.

The independent number

The independent anchor is Artificial Analysis, which ranks Carbon first of 57 models at 59 on version 5.0 of its Intelligence Index, released October 28. WinBuzzer's report on the launch says Claude Opus 5.5 sits at 58 and GPT-6 Astra at 57, with Argon at 50. That is a one-point lead over Anthropic, inside the range where run-to-run variation and settings matter. Carbon was measured at high reasoning effort, while the report says Opus and Astra ran at their maximum settings.

The cost of that lead is visible on the same page. Artificial Analysis shows Carbon using about 200 million output tokens on the index, against a median of 71 million, and a price of $7.76 per index task. WinBuzzer reports Argon at $7.68 per task, so a more capable model is not cheaper to run in practice.

Where Carbon loses

The model does not win everywhere. WinBuzzer's comparison has Carbon at 61.2% on FrontierSWE 2, behind GPT-6 Astra at 65.5% and Opus 5.5 at 62.3%. On the OSWorld offline subset, Carbon scores 81.5% against Opus 5.5 at 87.2%. On Artificial Analysis's Coding Agent Index, Carbon in Google's Antigravity tool scores 59 against 62 for Claude Code with Opus 5.5.

On CWE-bench, the same test Google leads with, the independent leaderboard tells a different story. With four attempts per task, Artificial Analysis lists Grok 4.7 at 81% and Opus 5.5 at 79%, and has no four-attempt Carbon result yet. Google's 71% is a single-attempt figure, where the report has Argon, GPT-6 Astra and Grok 4.7 at 68% and Opus 5.5 at 67%.

Carbon's factual reliability is mixed too. WinBuzzer reports an Artificial Analysis hallucination rate of 11%, up from Argon's 4%, though still among the lowest the site has recorded. We did not find that figure on the Artificial Analysis model page itself, so read it as a reported number.

Who can use it

Carbon is available first to a subset of cyber defenders through the Fairwind Program. The program's page says it works with over 650 partners and prioritizes governments and national cyber authorities, critical infrastructure operators and core technology platforms. Partners must pass background checks, use phishing-resistant multi-factor authentication, restrict access to internal security, incident-response or penetration-testing teams, and may not share or resell access.

Google gives no date for developers, businesses or consumers, saying only that broader availability will come as soon as possible. The precedent is Argon: announced for Fairwind partners on September 30, then opened to developers and Google AI Ultra and Pro subscribers on October 13, according to Google's Argon post. Google also says it is taking part in the US government's voluntary pre-release model testing.

Price

Google's introductory price is $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After an introductory period Google does not date, the rate becomes $4 and $20. Artificial Analysis currently lists the $4 and $20 rate. Argon's own promotional pricing ends November 30, 2026, per Google's Argon post.

Safety claims

Google says cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts at 0.4%, and that high-severity misalignment alerts in live coding environments fell by almost 80%. It reports 0.0% attack success on the Gray Swan indirect prompt-injection benchmark against 0.7% for Argon. These are internal results. For context on how Argon's safety numbers moved, see our look at Gemini 4 Argon's safety report.

What the index lead is worth

The index itself is new. Artificial Analysis rebuilt it as version 5.0 around mostly private tests and nine evaluations, and WinBuzzer reports Carbon tops four of the nine, including scientific research workflows and factual knowledge. A private test set is harder to game, but outsiders also cannot inspect it. History argues for caution about leads at this level: WinBuzzer notes Google's earlier index leads, Gemini 3 Pro in November 2025 and Gemini 3.1 Pro Preview in February 2026, were each overtaken within weeks. The same report says Google dropped a planned Gemini 3.5 Pro on September 30 in favor of Argon, according to a Google spokesperson speaking to Reuters. A one-point margin at the top of a fast-moving board is a snapshot, not a ranking you can plan around.

Our take

Carbon is a real step on an independent index, and a closely watched one. But a lead of one point at a different effort setting than its rivals is not a clear win, and the security claims that justify the restricted release are exactly the ones outsiders cannot yet reproduce. Watch for three things: a four-attempt CWE-bench entry from Artificial Analysis, the date Carbon leaves Fairwind, and whether the pre-release government testing Google mentions produces a public result.