Gemini 4 Carbon: Google’s New No. 1, Cyber Defenders First

Google has announced Gemini 4 Carbon, which it calls its “strongest model with state-of-the-art cybersecurity and long horizon capabilities.” The post, signed by Koray Kavukcuoglu, Google DeepMind’s SVP and Google’s chief AI architect, presents Carbon as the step after Gemini 4 Argon, with gains across coding, agentic work, science, full-stack web development and multimodal understanding. Most people cannot use it yet. Carbon goes first to a subset of cyber defenders, then to more trusted partners through Google DeepMind’s Fairwind Program, and only after that to developers, businesses and consumers, “as soon as possible.” Google gives no date for the wider release and does not mention the Gemini app, the Gemini API or Vertex AI.

The reason for the slow start is the model’s security skill. Google reports state-of-the-art scores on finding vulnerabilities (88.2% on SEC-bench Pro) and exploiting them (54.8% on ExploitGym), and 71.0% on CWE-bench, which tests patching. Internally, it says Carbon more than doubles Argon’s success rate at generating secure code, and that Chrome’s security team found it more than ten times better at producing high-quality patches than “another lab’s large frontier model,” which Google does not name. A model that is that good at exploitation is a hazard in the wrong hands, so defenders get the head start. Google also says it is taking part in the US government’s voluntary pre-release testing of frontier models, and it has published a safety and security report covering the Gemini 4 family.

The second pitch is endurance. Google says its own teams have watched Carbon work for days on multi-step engineering jobs, such as tracking down and fixing a bug in a large codebase, pushing through dead ends instead of stopping. In Google’s evaluation table, Carbon leads Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on Terminal-Bench 4.0 (65.2%), Agents’ Last Exam (77.4%), AutomationBench (53.9%) and Vending-Bench 2, where its simulated business ended with $16,484 against Astra’s $15,515. The clearest gap is long context: on GraphWalks, which tests reasoning over 256,000 to one million tokens, Carbon scores 96.2% to Argon’s 84.2% and Opus 5.5’s 66.8%. It does not win everything, though. Astra is ahead on FrontierSWE 2 (65.5% to 61.2%) and Opus 5.5 on OSWorld-2.1 (87.2% to 81.5%), and on WebDev Arena Carbon ranks fourth with 1,768 Elo, up 89 points.

Google also leans on behaviour. It says Carbon makes about half as many unsupported claims as Argon, asks for clarification instead of guessing, and flags discrepancies before acting. High-severity misalignment alerts in live coding sessions fell by almost 80%, overreach by 26% and under-service by 29%. On Gray Swan’s indirect prompt-injection benchmark, attacks succeeded 0.0% of the time against Argon’s 0.7%, and on Terminal-Bench 4.0 cheating attempts dropped to 0.4%, with none succeeding. That record is worth comparing with Argon’s first weeks, when Andon Labs caught it lying its way up the same Vending-Bench 2.

Independent numbers back the headline. Artificial Analysis puts Carbon first of 52 models on version 5.0 of its Intelligence Index, at 59, one point ahead of Opus 5.5 and two ahead of Astra, and Vals AI has it second on its own index at 70.4%. (Index v5.0 is a new scale; scores from BitsMinds’ earlier coverage on v4.3 are not comparable.) Artificial Analysis also measures what a finished task costs, which matters more than the price per token. Carbon is verbose, generating 200 million output tokens across the index against a median of 64 million, so at its full list price of $4 per million input tokens and $20 per million output tokens it costs $7.76 per task. That is cheaper than Opus 5.5, which has the same list price but burns more tokens, and dearer than Astra, which charges more per token but uses about a quarter as many:

Model (setting)	Intelligence Index v5.0	Cost per task	Output tokens for the index	List price per 1M tokens (in / out)
Gemini 4 Carbon (high)	59	$7.76	200M	$4 / $20
Claude Opus 5.5 (max)	58	$10.04	220M	$4 / $20
GPT-6 Astra (max)	57	$6.24	54M	$10 / $50
Gemini 4 Argon (high)	50	$7.68	110M	$4 / $20
Claude Fable 5.1 (max)	49	$13.04	160M	$10 / $50

Google’s own launch pricing is lower. For an introductory period Carbon costs $2 per million input tokens and $10 per million output tokens, the same as Argon, with cached input 95% cheaper; afterwards the price doubles to the $4 and $20 used above. Google has not said how long the introductory period lasts, and the Artificial Analysis page lists no public API yet. Its context window is one million tokens, and it accepts text, images, speech and video. For now, though, all of that matters mainly to the security teams Google lets in first. Everyone else gets a model with a strong benchmark sheet and no date for when they can try it, which leaves Anthropic and OpenAI a window to answer it.