Google Releases Gemini 4 Carbon, Surpassing GPT-6 Astra and Claude Opus 5.5

Google has launched Gemini 4 Carbon, a new frontier AI model that now ranks ahead of OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 on a major independent intelligence ranking. Carbon scores 59 on version 5.0 of the Artificial Analysis Intelligence Index, putting it in first place ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57).

Google announced Carbon as its strongest model yet, with major gains in long-running coding tasks, AI agents, and cybersecurity. It arrives just two weeks after Gemini 4 Argon, Google's previous flagship model.

There is a catch, though. Carbon is not open to the public yet. Google is first giving access to a small group of cybersecurity defenders and will expand it gradually before releasing it to developers, businesses, and consumers.

The New No. 1 AI Model

Google says Carbon shows significant gains in coding and agentic tasks. It also leads several benchmarks that matter for cybersecurity defense, including SEC-bench Pro (88.2%) for finding vulnerabilities, ExploitGym (54.8%) for exploiting them, and CWE-bench (71.0%) for patching them.

According to Google, Carbon also writes more secure code in its internal tests. In live coding work with Google's own teams, it showed lower rates of misalignment and made-up claims.

Built to Work for Days

Carbon is designed for hard, long-horizon engineering problems. Google says its teams have seen the model run for days, pushing past dead ends to complete complex, multi-step tasks such as tracing and fixing subtle bugs across large codebases and low-level systems.

Based on user feedback, Google improved Carbon across coding, agentic, and scientific tasks, full-stack web development, and multimodal understanding. Its web development output also improved, with Carbon gaining 89 Elo points to reach fourth place on WebDev Arena with a score of 1,768.

How It Stacks Up Against GPT-6 Astra and Claude

Google compared Carbon with Gemini 4 Argon, OpenAI's GPT-6 Astra, and Anthropic's Claude Opus 5.5 and Claude Fable 5.1. All results are Google's own.

Carbon leads most of the tests. It scores 65.2% on Terminal-Bench 4.0, well ahead of Claude Opus 5.5 (59.6%) and GPT-6 Astra (59.1%). On Agents' Last Exam v1.1, a computer-use test, it scores 77.4%, compared with 73.1% for GPT-6 Astra and 71.6% for Opus 5.5.

Its biggest lead is in long context. On GraphWalks, which tests reasoning across 256,000 to 1 million tokens, Carbon scores 96.2%. GPT-6 Astra scores 71.8% and Claude Opus 5.5 scores 66.8%. Carbon also tops AutomationBench (53.9%), RiemannBench for math (80.0%), and Vending-Bench 2, where it ended with a simulated bank balance of $16,484, compared with $15,515 for GPT-6 Astra.

Carbon does not win everywhere. GPT-6 Astra leads FrontierSWE 2 with 65.5%, ahead of Opus 5.5 (62.3%) and Carbon (61.2%). Claude Opus 5.5 leads OSWorld-2.1 with 87.2%, compared with 81.5% for Carbon. The Vals Index v2.2 is also nearly a tie, with Carbon at 70.4% and Opus 5.5 at 70.2%.

A Big Jump Over Gemini 4 Argon

Carbon is a clear step up from Gemini 4 Argon, which launched only two weeks earlier. It raises the Artificial Analysis Intelligence Index score from 50 to 59 and more than doubles Argon's result on the SWE-sweep coding test, from 4.3% to 9.6%.

Other gains are just as clear. Terminal-Bench 4.0 rises from 57.1% to 65.2%, GraphWalks from 84.2% to 96.2%, and ExploitGym from 42.0% to 54.8%. Google says its researchers and models are now working together to push capabilities forward faster.

A Cybersecurity Powerhouse

Google calls Carbon a significant leap in cybersecurity. It says the model is state-of-the-art across all three stages of cyber defense: finding vulnerabilities, checking whether they can be exploited, and fixing them.

On SEC-bench Pro, Carbon scores 88.2%, uncovering complex flaws hidden in real-world software on its own.

Defenders also use AI models to check which vulnerabilities can actually be exploited, so they know what to fix first. Carbon sets a new high score of 54.8% on ExploitGym, a popular exploitation benchmark.

For the final step, fixing the root cause, Carbon reaches 71.0% on Collinear's CWE-bench, turning confirmed findings into working code fixes.

Google says these external benchmarks were run without guardrails. In internal tests, Carbon was more than twice as successful as Argon at generating secure code. Google's Chrome team also reported that Carbon performed more than 10 times better than another lab's large frontier model at producing high-quality patches. Wiz, Google's cloud security platform, says Carbon leads its benchmarks for web penetration testing.

Harder to Trick, Less Likely to Cheat

Carbon uses the same layered safety system as Argon, which combines model training, real-time misuse defenses, prompt-injection protection, and chain-of-thought monitoring. Google says it went further on alignment this time and trained Carbon to handle complex or high-pressure tasks properly.

The company says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with cheating attempts falling to 0.4%. In live coding work with Google's internal teams, high-severity misalignment alerts dropped by almost 80%.

Carbon also makes about half as many unsupported claims as Argon in Google's internal testing. When information is unclear, it checks the relevant sources and points out the discrepancies to the user before acting. Google says it also reduced overreach by 26% and under-service by 29%, so it completes multi-step plans more reliably without going beyond its scope.

On Gray Swan's Indirect Prompt Injection Benchmark, Argon already had a low 0.7% attack success rate. Carbon brings that down to 0.0%, resisting every attack attempt in the test.

Price and Cost Per Task

Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens will cost 95% less than regular input tokens.

The introductory price is temporary. Once it ends, Carbon will cost $4 per million input tokens and $20 per million output tokens, double the launch rate. Google has not yet shared a cost-per-task figure for Carbon.

Who Can Use It and When

Google is taking a careful approach to the rollout. The company is taking part in the US government's voluntary pre-release model testing process.

Starting this week, a subset of cyber defenders will get access to Carbon. Google will then slowly add more trusted partners through its Fairwind Program. The company says it will keep collecting feedback and improving guardrails before bringing Carbon to developers, businesses, and consumers through its products as soon as possible.