Google unveils Gemini 4 Carbon, its top-scoring AI model, with access limited to cybersecurity defenders

Launch and benchmark standing

Google announced Gemini 4 Carbon on Wednesday, 28 October, calling it the most capable frontier model it has released so far. The launch came about two weeks after Gemini 4 Argon, which Google introduced on 30 September and opened more broadly on 13 October. Koray Kavukcuoglu, senior vice president and chief AI architect, leads the work at Google DeepMind.

Independent evaluator Artificial Analysis gives Carbon 59 points on its Intelligence Index v5.0, using high reasoning, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Argon at 50. Google's own table places Carbon first in 15 of 17 benchmark rows, while OpenAI wins one software engineering test and Anthropic another on computer use.

Gemini 4 Carbon59

Claude Opus 5.558

GPT-6 Astra57

Gemini 4 Argon50

Gemini 4 Carbon

59 points

Claude Opus 5.5

58 points

GPT-6 Astra

57 points

Gemini 4 Argon

50 points

Data from the article text

Outside testing found Carbon improves considerably on Argon in coding and knowledge work, yet it trails several rivals on professional tasks. The same testing recorded a higher hallucination rate than Argon on a factual-knowledge test, so the lead on Google's tables does not carry over everywhere.

Cybersecurity results

Cybersecurity is the centrepiece of the release. Google reports that Carbon scored 88.2% on SEC-bench Pro, which measures how well a model finds flaws in real software, 54.8% on ExploitGym, which tests whether a vulnerability can be turned into an exploit, and 71% on CWE-bench, which covers fixing security problems by changing code. Google says those runs were made without the model's usual protective mechanisms and do not show it can compromise every application or infrastructure. The company also claims Carbon produces secure code at more than twice the success rate of Argon.

SEC-bench Pro88.2%

ExploitGym54.8%

CWE-bench71%

SEC-bench Pro

88.2 %

ExploitGym

54.8 %

CWE-bench

71 %

Data from the article text

In the Grey Swan evaluation, a prompt-injection test built around hidden instructions, Carbon recorded a 0% success rate for the attacks tested. Argon scored 0.7%, Claude Opus 5.5 1% and GPT-6 Astra 8.5%. That figure is not presented as a guarantee against every possible attack.

Gemini 4 Carbon0%

Gemini 4 Argon0.7%

Claude Opus 5.51%

GPT-6 Astra8.5%

Gemini 4 Carbon

0 %

Gemini 4 Argon

0.7 %

Claude Opus 5.5

1 %

GPT-6 Astra

8.5 %

Data from the article text

Long-running coding and agent work

Google says its internal teams have run Carbon for several days in a row on complex, multi-step engineering work. The model reportedly tests alternative paths after dead ends until it finds the root cause of hidden faults in large codebases. On the WebDev Arena leaderboard it gained 89 Elo points, reaching 1,768 and placing fourth overall.

Introductory pricing matches the level Argon launched at, $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. A footnote sets the later rates at $4 and $20, and Google has not said when the introductory period will end.

Restricted rollout

Access is limited for now. Carbon is available only to a selected group of cybersecurity defenders, and Google plans to extend it to trusted partners through the Fairwind program as it evaluates safeguards against misuse. The company is also taking part in a voluntary pre-release testing framework run with US authorities.

Google intends to bring Carbon to developers, businesses and everyday users, but it has not named a date. The company says a broader launch will follow once it has collected feedback and refined its guardrails. Argon stays the model customers can use under the existing public offer.