Google's Gemini 4 Carbon Takes #1 on Artificial Analysis, Opens Only to Cyber Defenders

TL;DR

Google announced Gemini 4 Carbon on October 28 and calls it its strongest model, with the focus on cybersecurity and long-horizon work. Artificial Analysis puts it first on its Intelligence Index v5 at 59, nine points above Gemini 4 Argon. Its cost per index task, $7.76, sits between GPT-6 Astra and Claude Opus 5.5. You can't call it yet. Access starts with a subset of cyber defenders in Google's Fairwind Program, and developers, businesses and consumers get it "as soon as possible." The price is already set: $2 input and $10 output per million tokens during the introductory period, then $4 and $20. Independent tests also found a cost Google's post leaves out. The hallucination rate went from 4% on Argon to 11%.

What Google shipped, and to whom

Carbon arrives the same way Argon did. A subset of cyber defenders gets it this week, and access widens to more trusted partners through Fairwind. Google says it will open Carbon to everyone else after it collects feedback and tightens guardrails. The announcement mentions no API endpoint, no Gemini app rollout and no date.

Pricing, at least, is public:

• Introductory: $2 per million input tokens and $10 per million output tokens, the same as Argon's introductory rate.
• After the introductory period: $4 input and $20 output.
• Cached input: 95% off the input price, which comes to $0.20 per million at the standard rate according to Artificial Analysis.
• Context: 1M tokens, according to Artificial Analysis. Latent Space also lists 1M max output, with text, image, video and speech in and text out.

Google gives no end date for the introductory rate. Artificial Analysis describes the same arrangement as $4/$20 list with a 50% launch discount. If you budget for this model, budget at $4/$20 and count the discount as a bonus.

The cyber numbers

Cybersecurity is Google's headline, and it reports three results run without guardrails: 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 calls the first two state of the art. It also claims more than 2x Argon's success rate at secure code generation on an internal test.

Partners back some of this up. Google says the Chrome team found Carbon more than 10x better than "another lab's large frontier model" at writing high-quality patches. It does not name the model. Wiz reports Carbon leading its web penetration testing benchmarks. On Gray Swan's indirect prompt injection benchmark, Carbon's attack success rate was 0.0%, down from Argon's 0.7%.

That set of results explains the access policy. A model that finds and exploits bugs this well is useful to an attacker too, so Google gives it to the people who patch first. Opening it to everyone stays on the to-do list with no date.

What the independent evals say

Artificial Analysis tested Carbon at high reasoning effort. Its score of 59 is the highest on Intelligence Index v5, and Latent Space's roundup of the AA results says Carbon leads four of the index's nine evals. The rest of the scorecard is uneven:

• Coding agents: Carbon scores 59 on AA's Coding Agent Index in Google's Antigravity CLI. Claude Opus 5.5 at max effort scores 62 in Claude Code.
• Terminal work: 64% on Terminal-Bench 4.0, ahead of Opus 5.5's 56%.
• FrontierCode v1.1: 51%, behind Opus 5.5 at 56%.
• AA-Briefcase (knowledge work): 1609 Elo, up from Argon's 1483, but behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro, which leads at 1678.

Cost is where Carbon looks good. Running one index task costs $7.76 at standard pricing, against $10.04 for Opus 5.5 at max effort and $6.25 for GPT-6 Astra at max. Carbon used 123k output tokens per task to Opus's 173k.

The number Google left out

Google's post says Carbon makes about half as many unsupported claims as Argon. AA-Omniscience measures something else: how often a model answers a question it should decline. On that test Carbon's hallucination rate is 11%, up from Argon's 4%. Latent Space's roundup of the AA data says 11% is still the second-lowest rate measured. Opus 5.5 and GPT-6 Astra both sit at 50%.

Picture two contractors. One says "I don't know" when unsure. The other gives an answer every time, and some of those answers are made up. Argon was the first contractor. Carbon still mostly is, but it tries to answer more often. In vulnerability triage, a confident report about a bug that doesn't exist costs an engineer an afternoon, and Latent Space's roundup quotes users raising that worry.

Safety claims

Google reports these from internal testing. Successful cheating on Terminal-Bench 4.0 fell to 0.0%, and cheating attempts fell to 0.4%. High-severity misalignment alerts in live coding environments dropped almost 80%. Overreach fell 26% and under-service fell 29%. Carbon runs on Argon's safety stack, which covers training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. Google says it is also taking part in the US government's voluntary pre-release testing. The details are in a separate safety report.

Google also says internal teams have watched Carbon work for days on complex engineering tasks. It gives no benchmark for that, so treat it as an anecdote until outside testers can run long jobs themselves.

What to do with this if you build

For most of you, nothing yet. You can't call Carbon unless you are a vetted defender. The results still matter for planning: Google now holds the top spot on AA's index, and that model costs less per task than Opus 5.5 at max.

• Price it at $4/$20. The introductory rate has no end date, and you won't get it until public access opens anyway.
• Don't swap your coding agent yet. Opus 5.5 still leads AA's Coding Agent Index and FrontierCode, and Carbon's coding numbers come from one harness, Antigravity CLI.
• Re-test refusals. If your pipeline relied on Argon declining when unsure, the jump from 4% to 11% is the regression to check first.
• Watch Fairwind. Argon started in the same program. Google says Carbon access widens to more trusted partners there before anyone else gets it.

Key Takeaways

• Gemini 4 Carbon scores 59 on Artificial Analysis Intelligence Index v5, first place and nine points above Gemini 4 Argon.
• Only a subset of cyber defenders in Google's Fairwind Program can use it now. Google gives no date for developers, businesses or consumers.
• Pricing is $2/$10 per million tokens introductory and $4/$20 after, with cached input 95% off and a 1M-token context.
• Google-reported cyber results: 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, all without guardrails.
• Carbon costs $7.76 per index task against $10.04 for Opus 5.5 at max, but trails Opus on AA's Coding Agent Index (59 vs 62) and FrontierCode (51% vs 56%).
• Its AA-Omniscience hallucination rate rose from Argon's 4% to 11%, still well below the 50% measured for Opus 5.5 and GPT-6 Astra.

Sources: Google, Introducing Gemini 4 Carbon, Artificial Analysis, Gemini 4 Carbon is the new #1, Latent Space AINews, Gemini 4 Carbon tops leaderboards, Google DeepMind Fairwind Program, Gemini 4 safety report, Artificial Analysis Intelligence Index