Google has introduced Gemini 4 Carbon, which it describes as its next frontier artificial intelligence model.
The model is being released first to a subset of cyber defenders. Access will expand gradually through Google's Fairwind programme before any wider launch to developers, businesses and consumers.
Google calls Gemini 4 Carbon its strongest model to date and is positioning the release around software engineering, coding and cyber defence. It says the system improves on Gemini 4 Argon, launched two weeks earlier, and that it is taking a deliberate approach to deployment while participating in the US government's voluntary pre-release model testing process.
Pricing matches the introductory level for Argon: USD $2 per million input tokens and USD $10 per million output tokens. Cached input tokens will cost 95% less than the standard input token rate.
Cyber focus
A large part of the announcement centred on cybersecurity. Google said Carbon led several external benchmarks covering different stages of cyber defence, including vulnerability detection, exploitation testing and patching.
According to Google, the model scored 88.2% on SEC-bench Pro for identifying vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. It also topped version 5.0 of the Artificial Analysis Intelligence Index with a score of 59.
Google said internal testing showed more than a twofold improvement over Argon in success rates for generating secure code. It also cited feedback from its Chrome team, which found Carbon performed more than 10 times better than another laboratory's large frontier model when generating patches.
Google also referenced Wiz, which it described as a cloud security platform, as observing Carbon leading benchmarks for web penetration testing. According to Google, that area is a core part of Wiz's Scan for Good initiative.
Longer tasks
Beyond cyber work, Google said Carbon was built for more difficult and extended engineering tasks. Internal teams, it said, had seen it operate for days while working through dead ends on multi-step jobs such as tracing subtle bugs across large codebases and low-level systems, and producing patches.
Google also said it had focused on coding, agentic and scientific tasks, as well as full-stack web development and multimodal understanding. On WebDev Arena, it said Carbon gained 89 Elo points to reach fourth place with a score of 1768.
The release reflects a broader pattern in the AI market, with model developers trying to show progress not only through general reasoning tests but also through narrower measures in coding, security and practical software work. For Google, the emphasis on cyber defence may also help distinguish Carbon in a crowded field of frontier models, where benchmark gains alone can be difficult for customers to assess.
Safety measures
Google also used the launch to outline changes it said were intended to reduce harmful or unreliable behaviour. Carbon builds on the safety and security stack used in Argon, it said, including model training, real-time misuse defences, prompt-injection resilience and chain-of-thought monitoring.
Among the measures highlighted were lower rates of unsupported claims, reduced overreach and under-service in multi-step execution, and stronger resistance to prompt injection. Google said Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, while attempts to cheat fell to 0.4%.
It also said that in live coding environments with internal teams, high-severity misalignment alerts dropped by almost 80%. In internal testing, unsupported claims were cut by about half compared with Argon.
Google reported that Carbon also achieved a 0.0% attack success rate on Gray Swan's Indirect Prompt Injection benchmark, compared with Argon's previously reported 0.7%. It said the model was trained to respond appropriately to complex or high-pressure tasks and to surface uncertainties to users rather than guessing or acting beyond scope.
"Today, we're announcing the next frontier model: Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities," said Koray Kavukcuoglu, SVP, Google DeepMind and Chief AI Architect, Google.
"We're taking a deliberate approach to rollout and are actively engaged in the U.S. government's voluntary process for pre-release model testing," said Kavukcuoglu.