Gemini 4 Carbon tops the Intelligence Index, but only cyber defenders can use it yet

What Google announced

On October 28, 2026, Google announced Gemini 4 Carbon, “our strongest model with state-of-the-art cybersecurity and long horizon capabilities.” It comes two weeks after Gemini 4 Argon, which Google released on October 13.

Google leads with four claims:

• First on Artificial Analysis’s Intelligence Index v5, at 59. On the same leaderboard, Claude Opus 5.5 at max effort scores 58 and GPT-6 Astra 57.
• Cybersecurity. Carbon scores 88.2% on SEC-bench Pro, for finding vulnerabilities, 54.8% on ExploitGym, for exploiting them, and 71.0% on CWE-bench, for patching them. Google ran those tests “without guardrails.”
• Long tasks. Carbon is “built for sustained agentic workflows,” and Google says its teams “have seen Carbon run for days to persist through dead ends” on work such as “root-causing and patching subtle bugs across large codebases.”
• Web development. It “jumps +89 elo points to land at #4 (1769)” on WebDev Arena.

Who can use it, and when

For now, very few people. Google’s rollout, in its words: “Starting this week, we will make Carbon available to a subset of cyber defenders and will gradually phase access to additional trusted partners through our Fairwind Program. We’re continuing to gather feedback and iterate on guardrails before making Carbon available to developers, businesses, and consumers via our product surfaces as soon as possible.” It adds that it is “actively engaged in the U.S. government’s voluntary process for pre-release model testing.”

The Fairwind Program gives “high-priority defenders (like governments, healthcare providers, and telecommunications services) early access to advanced models.” It has “over 650 partners,” runs background checks on organizations that apply, and lets members give access only to “internal cybersecurity, incident response, or penetration testing teams.” On October 28 its page still described Argon as the model it offers.

Here is where Carbon is, beside Argon, as Google’s pages read on October 28:

Where	Gemini 4 Carbon	Gemini 4 Argon
Gemini app	Not yet	Google AI Pro ($19.99 a month) and Ultra
Google Antigravity	Not yet	Paid Google AI Pro, Ultra and Enterprise, in the app and CLI
Gemini API and AI Studio	Not yet	Paid tier only
Fairwind Program	A subset of cyber defenders, from this week	Some members, in a version for cyber defense

The sources are the Gemini API’s model list, Antigravity’s models page, Google’s Antigravity post on Argon and Google’s AI plans. Artificial Analysis has tested Carbon, but its Coding Agent Index marks the model as unavailable.

How it compares

Carbon against the models you can use, on Artificial Analysis’s Intelligence Index v5, with the cost of one index task at API prices:

Model	Index	Cost per task	API price per million tokens, in and out	Where you can use it
Gemini 4 Carbon (high)	59	$7.76	$2 and $10 at launch, then $4 and $20	Fairwind Program only
Claude Opus 5.5 (max)	58	$10.04	$4 and $20	Claude Pro ($20 a month) and Max
GPT-6 Astra (max)	57	$6.24	$10 and $50	ChatGPT Plus ($20 a month) and Pro
GPT-6.1 Sol (max)	55	$1.27	$2 and $10	ChatGPT Plus and Pro, in Work and Codex
Gemini 4 Argon (high)	50	$7.68	$2 and $10 to November 30, then $4 and $20	Google AI Pro and Ultra

Artificial Analysis prices both Gemini 4 models at $4 and $20, the price after Google’s introductory period, so at launch prices a task would cost less. The plans are from Claude’s pricing, ChatGPT’s pricing and Google’s; Argon’s dates are on the Gemini API pricing page.

The lead is one point, and it isn’t everywhere. Of the index’s nine tests, Carbon leads four: HLE-Diamond (69%), Terminal-Bench-Science (65%), AutomationBench-AA (78%) and AA-Omniscience, which rewards correct answers and penalizes made-up ones. Opus 5.5 still leads the coding test, FrontierCode 2.0, at 46% to Carbon’s 41%, and Kimi K3.1 Pro leads Terminal-Bench 4.0, at 67% to Carbon’s 65%.

Google’s own table compares Carbon with Argon, GPT-6 Astra, Opus 5.5 and Claude Fable 5.1 on 17 results, and Carbon leads 15. The two it doesn’t: GPT-6 Astra leads FrontierSWE 2, a coding test, at 65.5% to Carbon’s 61.2%, and Opus 5.5 leads OSWorld-2.1, computer use, at 87.2% to 81.5%. Vals AI, testing on its own, ranks Carbon second of 29 models on its Vals Index, at 70.36%.

For coding

Artificial Analysis also ran Carbon in Google’s Antigravity CLI for its Coding Agent Index, which scores an agent and its model together. Here it is beside the setups that lead the index, and beside Argon in the same agent, on October 28:

Agent and model	Index	Cost per task	Time per task
Claude Code, Opus 5.5 (max)	61.7	$11.42	61 min
Devin Fusion CLI, Opus 5.5 with GPT-6 Luna	61.5	$8.00	51 min
GitHub Copilot, Copilot Fusion (Max)	59.9	$5.59	36 min
Antigravity CLI, Gemini 4 Carbon (high)	59.4	$7.74	30 min
Claude Code, Sonnet 5.5 (max)	59.3	$13.30	80 min
Antigravity CLI, Gemini 4 Argon (high)	53.2	$11.73	35 min

In the same agent, Carbon scores 6.2 points more than Argon, for about two-thirds of the cost per task and in less time. Only Claude Code with Opus 5.5 scores higher with a single model. Devin Fusion CLI and Copilot Fusion, which GitHub released the same day, each combine more than one model.

What Google says about safety

Google’s announcement makes these claims for Carbon, from its own testing:

• Less cheating. It “reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4%.”
• Fewer alarms. “In live coding environments with our internal teams, we observed that high-severity misalignment alerts dropped by almost 80%.”
• Fewer made-up claims. It “cuts unsupported claims by approximately 2x over Argon.”
• Staying in scope. It “reduced overreach by 26% and under-service by 29%.”
• Prompt injection. On Gray Swan’s indirect prompt injection benchmark, attacks succeeded 0.0% of the time, against 0.7% for Argon.

The gate is about the other side of its cyber skills. As the Fairwind page puts it, “in the wrong hands, the same capabilities can become an equally powerful threat.”

What to do now

1. If you defend critical systems, apply to Fairwind. Google puts “governments and national cyber authorities,” “critical infrastructure operators” and “core technology platforms” first, and vets every applicant.
2. Otherwise, choose from what you can use. Claude Opus 5.5 is a point behind Carbon on the index and leads its coding test; GPT-6.1 Sol scores 55 for about an eighth of Opus’s cost per task; and Argon is Google’s best on Google AI Pro. For coding agents, Claude Code with Opus 5.5 still leads.
3. Don’t buy a Google plan for Carbon yet. Google hasn’t said which plans will get it. Argon reached Antigravity on Google AI Pro and Ultra, but Google hasn’t said Carbon will follow.
4. If you build on the API, budget for the full price. Google says Carbon’s price rises to $4 and $20 “after the introductory period expires.” Argon’s rises on December 1, 2026.

What Google hasn’t said

• When. Beyond “as soon as possible” and “soon,” there’s no date for developers, businesses or consumers.
• Which plans. Nothing yet on Google AI Pro or Ultra, the Gemini app, Antigravity or Jules.
• How long the launch price lasts. Argon’s runs through November 30, 2026; Carbon’s end date isn’t given.
• The details developers need. There’s no model ID or Carbon page in the Gemini API docs yet. Artificial Analysis and Vals AI list a context window of 1 million tokens.
• A safety report of its own. The announcement links the Safety and Security report Google published with Argon, not one for Carbon.

When it reaches Antigravity

Enjoy runs Antigravity on your Google account, beside Claude Code, Codex and the other labs’ agents, each on its own plan. Enjoy asks an agent for its current models each time you open its settings, so when Google adds Carbon to Antigravity for your plan, it appears in the model list without an Enjoy update. Until then, you can give the hardest work to Claude Code with Opus 5.5, and switch agents in the middle of a conversation when you want Argon’s view on it.

Enjoy is a monthly subscription, with the first seven days free (pricing). It doesn’t add Carbon to an account Google hasn’t given it to.

We’ll update this memo when Google opens Carbon to developers or adds it to its plans.