Google’s Gemini 4 Carbon looks like a diamond in the rough
Google has announced Gemini 4 Carbon, a stronger version of its Argon model, just two weeks after Argon’s launch. Google says Carbon is particularly capable at cybersecurity and long-running tasks, and its early benchmark results strengthen the company’s position against Anthropic and OpenAI.
Carbon initially costs $2 per million input tokens and $10 per million output tokens, matching Argon’s price for an unspecified introductory period; prices are then due to rise to $4 and $20, respectively, matching Claude Opus 5.5. Carbon leads several cybersecurity benchmarks and ranks fourth on WebDev Arena, while Google says it more than doubled Argon’s secure-code success rate on internal tests. The model will be available this week to selected cybersecurity defenders, with wider access planned.
In short
- Google has launched its stronger Gemini 4 Carbon model.
- Introductory prices match Argon before doubling.
- Selected cybersecurity defenders will get access this week.
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Carbon’s early benchmark results and reported gains in secure coding suggest Google has made a meaningful advance, especially for defenders handling complex, long-running cybersecurity work. Its introductory pricing matches Argon’s, and wider access could give more users a capable alternative to leading models.
The case against
The evidence is still preliminary: some of the security claims come from Google’s own internal tests, and access is initially limited, making independent assessment difficult. Carbon’s introductory price is temporary and is set to rise substantially, while its fourth-place WebDev Arena ranking suggests its strengths may not extend equally across tasks.
Explained simply
Google’s Gemini is a family of artificial intelligence models made by Google. These systems process text and other information, and can be used for tasks such as writing code or carrying out multi-step work. Google’s Argon model launched recently, and Carbon is a newer, stronger version.
Google is competing with firms including Anthropic, which makes the Claude models, and OpenAI. Researchers compare models on tests, or benchmarks, but results can vary depending on what each test measures. Cybersecurity tests assess how models handle tasks related to finding or fixing security weaknesses, while coding rankings measure performance on programming tasks.
AI models are often charged for the amount of text they process and produce, counted in tokens, so prices can affect who can afford to use them. Access to powerful models for cybersecurity can also matter because they may help defenders identify weaknesses, while the same abilities could be used in attacks.
AI Business Climate Companies Science Technology