Google's Gemini 4 Carbon looks like a diamond in the rough

Just two weeks after unveiling Gemini 4 Argon, Google has introduced an enhanced iteration dubbed Gemini 4 Carbon. This model is set to be sold at the same price point as its predecessor, at least for a limited time. Koray Kavukcuoglu, Senior Vice President at Google DeepMind and chief AI architect at Google, declared that Gemini 4 Carbon is the company's most formidable model to date, especially in terms of cybersecurity and prolonged tasks.

The moniker 'Carbon,' despite its apparent irony, reflects Google's commitment to a carbon-neutral future. The name could have been more fitting, such as 'Gemini 4 Silicon,' but 'Si' is not as catchy.

Priced initially at $2 per million input tokens and $10 per million output tokens, Gemini 4 Carbon matches the cost of its predecessor Gemmi 4 Argon during an indefinite introductory period. Following the introduction phase, Google plans to raise prices to $4 per million input tokens and $20 per million output tokens, positioning the model competitively against Anthropic's Claude Opus 5.5, which Google anticipates as Carbon's main rival.

Google's capability to thrive following the leadership change at its DeepMind AI division appears promising, thanks to Gemini 4 Carbon's impressive benchmark scores. In tests, it excels in SEC-bench Pro, ExploitGym, and CWE-bench, ranking fourth in the WebDev Arena leaderboard behind Opus 5.5, GPT-6.1 Astra, and Sonnet 5.5. Currently, it stands second to Opus 5.5 in Arena.ai's Best Overall Agent ranking.

Review site Artificial Analysis states that Gemini 4 Carbon outperforms Google's index and ranks second in its Coding Agent Index, behind Claude Code Opus 5.5.

Kavukcuoglu highlighted the model's cybersecurity capabilities, accessible to a select group of cybersecurity experts this week and eventually through Google's Fairwind access program. Internal benchmarks reveal that Gemini 4 Carbon achieves a significant 2x improvement in successful secure code generation compared to Argon and a staggering 10x enhancement over another large frontier model when generating high-quality patches.

According to Google's CEO, Sundar Pichai, Gemini 4 Argon and Carbon are currently leading in the frontier of AI models. Google's first-party model APIs have processed nearly 32 billion tokens per minute, a fourfold increase from a year ago. Despite Google's aspiration to surpass Anthropic and OpenAI in model performance and usage, the company already leads in revenue.

Google boasts a half-trillion-dollar run rate, nearly ten times more than either OpenAI or Anthropic.