Google launches Gemini 4 Carbon, now ranked No. 1 by Artificial Analysis

Google announces Gemini 4 Carbon on 28 October, two weeks after releasing Gemini 4 Argon. It calls Carbon its “strongest model with state-of-the-art cybersecurity and long horizon capabilities”. The headline fact: Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. That puts it one point ahead of Anthropic’s Claude Opus 5.5, which led the rebuilt index at launch with 58. Earlier coverage of that rebuild is in our report on Artificial Analysis Index v5.

Builders cannot use it yet. Google is releasing Carbon first to a subset of cyber defenders this week. Access then widens gradually to trusted partners through its Fairwind Program, before Carbon reaches developers, businesses and consumers.

What Google announced on 28 October

Koray Kavukcuoglu, senior vice president of Google DeepMind and Chief AI Architect at Google, wrote the launch post. Google says Carbon shows “significant gains across coding and agentic tasks” and in capabilities “critical for cybersecurity defense”. Based on user feedback, it focused on coding, agentic and scientific tasks, plus full-stack web development and multimodal understanding.

Google CEO Sundar Pichai promoted the release on X:

“Following Argon’s release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Intelligence Index at 59.”

Pichai’s post appears in Seeking Alpha’s report, which also says Alphabet shares rose about 1.5% in early afternoon trading after the announcement.

Carbon was reported before launch. On 10 October, a 4SAPI post summarised a Business Insider report saying Google was testing an internal build codenamed “Carbon,” built on the Argon architecture. The report cited internal feedback that its coding might approach Opus 5.5. Google declined to comment at the time. Google’s launch post now confirms the model, but not those internal details.

Spec	Gemini 4 Carbon	Source basis
Artificial Analysis Intelligence Index v5	No. 1, score 59	Artificial Analysis, cited by Google
Context window	1 million tokens	Reported via Artificial Analysis
Inputs / output	Text, images, video, speech / text	Reported via Artificial Analysis
Introductory price	$2 input, $10 output per million tokens; cached input 95% off	Google
Price after introduction	$4 input, $20 output per million tokens	9to5Google
Access	Cyber defenders, then Fairwind partners, then general	Google

Where Carbon leads and trails on Artificial Analysis

Artificial Analysis is an independent testing service. Its Intelligence Index combines many evaluations into one number. Trending Topics reports the index’s detailed results. Carbon has the top score in four of the nine individual tests: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. That puts it ahead in complex reasoning, knowledge and scientific tasks. It gains 9 points over Argon.

The lead is narrow, and Opus 5.5 still wins on agentic coding:

• Coding Agent Index: Carbon scores 59 running in the Antigravity CLI. Claude Code with Opus 5.5 scores 62.
• Terminal-Bench 4.0 and FrontierCode v1.1: Carbon scores 64% and 51% and ranks second behind Opus 5.5 on both.
• AA-Briefcase (an agentic knowledge-work test): Carbon reaches 1609 Elo, up 126 points from Argon. It still trails Grok 4.7 (1638), Muse Wonder (1659) and Kimi K3.1 Pro (1678). For more on the leader there, see our Kimi K3.1 Pro report.

On reliability, Carbon scores 54 on the updated AA-Omniscience test, against 48 for Opus 5.5 and 46 for Argon. Its accuracy rises 10 points over Argon to 58%. Its hallucination rate, meaning how often it states false information, rises from 4% to 11%. Artificial Analysis says that is still the second-lowest of all models, behind only Argon. Opus 5.5 and GPT-6 Astra both come in at 50%.

The index carries its own caveats. Version 5 was only recently rebuilt with secret test data. Shattered’s analysis notes that composite scores from different index versions cannot be compared. It adds that a model can lead the average while trailing on a specific task a team cares about.

Cybersecurity claims: find, exploit, patch

Google pitches Carbon mainly as a cyber-defence tool. It reports one external benchmark for each stage of a vulnerability’s life, all evaluated without guardrails:

• Finding flaws: 88.2% on SEC-bench Pro, which Google calls state of the art. Seeking Alpha notes this beats the 85.4% scored by OpenAI’s GPT-6 Astra.
• Exploitation: 54.8% on ExploitGym. Google says defenders increasingly use models to check which vulnerabilities can actually be exploited, so they can prioritise fixes.
• Patching: 71.0% on Collinear’s CWE-bench, which tests turning validated findings into working code fixes.

Google adds several internal results. Carbon achieves “more than 2x uplift in success rate for generating secure code over Argon.” Google’s Chrome team reported that Carbon “performed more than 10X better than another lab’s large frontier model when generating high quality patches”. Google does not name that lab. It also says the cloud security platform Wiz sees Carbon leading on web penetration-testing benchmarks. These results are company-reported. Trending Topics stresses that the figures come from Google.

Long-horizon coding and web development

“Long-horizon” here means tasks that take many steps and a long time to finish. Google says Carbon is “built for sustained agentic workflows.” Its internal teams have seen it “run for days to persist through dead ends,” for example to find the root cause of subtle bugs and patch them across large codebases and low-level systems. Pichai said the model already shows improvement over Argon on long-horizon tasks.

On WebDev Arena, a leaderboard for web development, Google says Carbon gains 89 Elo points and lands at No. 4. Google’s post gives its score as 1768. 9to5Google and Trending Topics both report 1769.

Artificial Analysis also measured cost to complete each task in the index. At standard prices, Carbon costs $7.76 per task. That is roughly 23% less than Claude Opus 5.5 ($10.04), but 1.2 times the cost of GPT-6 Astra ($6.25). Carbon uses 123,000 output tokens per task, 29% fewer than Opus 5.5’s 173,000. For teams running long agent jobs, token use often matters as much as the list price.

Safety, alignment and the staged rollout

Carbon uses the same multi-layered safety stack as Argon: model training, real-time misuse defences, resistance to prompt injection, and chain-of-thought monitoring. Prompt injection means hidden instructions slipped into content a model reads. Chain-of-thought monitoring means checking the model’s step-by-step reasoning for problems. Google reports these behaviour metrics, all from its own testing:

• Cheating on Terminal-Bench 4.0: successful cheating falls to 0.0%, and cheating attempts fall to 0.4%, while overall accuracy improves.
• Misalignment alerts: high-severity alerts, which flag the model acting against its instructions, dropped by almost 80% in live coding work with internal teams.
• Honesty: about half as many unsupported claims as Argon.
• Staying in scope: overreach down 26% and under-service down 29%.

Trending Topics adds that Carbon was not successfully attacked in any case in Gray Swan’s prompt-injection test, again per Google.

Google calls its rollout “deliberate.” It says it is taking part in the US government’s voluntary process for testing models before release. Fairwind is the same channel Google used for Argon. Argon was announced on 30 September to Fairwind cyber defenders and opened to paid users on 13 October. On that date it became generally available on the Gemini API and reached Google AI Pro and Ultra subscribers in the Gemini app.

What Carbon means for builders

For now, Carbon is a model to plan for, not one to build on. Google’s public line is that it is “coming soon to developers, businesses, and consumers”, with no date. Shattered warns that anyone planning a migration around Carbon is “planning around a promise, not a ship date”.

Pricing is clearer. Google sets Carbon’s introductory price equal to Argon’s: $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. Cached input means repeated prompt content the provider has stored and charges less to reuse. 9to5Google reports the price rises to $4 and $20 afterwards. Artificial Analysis says no end date for the discount has been confirmed. For comparison, Argon’s introductory pricing was reported to run through 30 November.

The pattern for teams choosing models is a split. Carbon leads on reasoning, knowledge and token efficiency in independent testing, but Opus 5.5 keeps the edge on agentic coding. Teams focused on coding agents should test both on their own repositories before switching. Teams doing security work have the most reason to watch Fairwind access. Every security figure except the GPT-6 Astra comparison comes from Google itself.

What to watch next

• A general release date and any model card or safety report for Carbon. Google published both for Argon.
• Independent cybersecurity tests that check Google’s SEC-bench Pro, ExploitGym and CWE-bench results.
• Anthropic’s response. Trending Topics speculates that Anthropic could unveil a new top model before its IPO, with no confirmation from the company. Background is in our report on Anthropic’s filing.
• Head-to-head tests against OpenAI’s newest models, including GPT-6.1 Terra. Shattered notes that no independent comparison of Carbon against those models exists yet.