Gemini 4 Carbon: Our next frontier model
Google ● Covered by 2 sources
Google DeepMind unveiled Gemini 4 Carbon, a new model aimed at coding and cybersecurity. It’s rolling out slowly after strong benchmark results, including a first-place rank on Artificial Analysis.
Based on reporting by Google — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind has unveiled Gemini 4 Carbon, a new frontier model it says is its strongest yet. The pitch is clear: better coding, better long-horizon work, and a bigger jump in cybersecurity than the company’s last release, Gemini 4 Argon, which arrived two weeks ago.
Carbon sits at the top of the recently updated v5.0 Artificial Analysis Intelligence Index with a score of 59. The company says it also leads on several cybersecurity benchmarks, including SEC-bench Pro at 88.2% for finding vulnerabilities, ExploitGym at 54.8% for exploitation, and CWE-bench at 71.0% for patching. Internally, Google says the model produces more secure code and shows fewer misalignment and fabrication issues in live coding work with its teams.
The more interesting claim is not that Carbon is fast, but that it can keep going. Google says it was designed for sustained agentic workflows and has run for days internally while teams used it on multi-step engineering problems, including rooting out subtle bugs across large codebases and low-level systems. It also says the model improves on a range of areas that matter to builders, from full-stack web development to multimodal understanding.
There is a price tag, too. Carbon will launch at the same introductory pricing as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input token price. But access will stay tight at first. Google says it is working through the U.S. government’s voluntary pre-release testing process, and that this week Carbon will go to a subset of cyber defenders before expanding to more trusted partners through its Fairwind Program.
Safety is a big part of the sales pitch here, not an afterthought. Google says Carbon uses the same multilayered safety stack as Argon, with training, misuse defenses, prompt-injection resilience, and chain-of-thought monitoring. It also says the model reduced successful cheating on Terminal-Bench 4.0 to 0.0%, cut unsupported claims by about 2x over Argon, and resisted every evaluated prompt-injection attack in one benchmark. The company wants the model to look powerful without looking reckless, which is basically the whole story of modern frontier AI.
My take — AI-written commentary, not fact-checked reporting
This is the usual frontier-model move: post the benchmark numbers first, then promise the guardrails are keeping pace. The real tell is the slow rollout to cyber defenders and trusted partners, because when a model is this good at code and exploits, nobody wants to be the first public customer. Google is selling competence with a seatbelt, and that’s probably the right order.
Read more about this at: Google