Google has unveiled Gemini 4 Carbon, just two weeks after opening Gemini 4 Argon to paying Gemini subscribers and developers. According to the Mountain View company, this new model, the most powerful in its lineup, takes the top spot on the Artificial Analysis Intelligence Index, ahead of Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra. Its access remains limited to a select circle for now.
Long Tasks and Comprehensive Cyberdefense
Google DeepMind designed Gemini 4 Carbon for prolonged agentic workflows, meaning sequences of actions carried out by the model itself, on complex engineering problems that unfold in many steps. In its blog post, the company claims to have seen the model work for several days on the same task, persevering when it hits a dead end. Its results also improve in “full-stack web development, multimodal understanding, and on scientific tasks, ” according to Google.
In the field of cybersecurity, the firm asserts that Carbon now covers all stages of system defense: vulnerability detection, including complex flaws in existing software; validation, which involves verifying that a flaw can be exploited to know which ones to fix as a priority; and remediation, with code that addresses the root cause of the identified flaw. Google also highlights a clear improvement in secure code generation compared to Argon.
Safeguards Inherited from Argon and a More Cautious Model
Carbon retains the protections implemented for Argon, from real-time detection of malicious uses to monitoring its chain of thought, meaning the intermediate reasoning it produces before its response. Google says it went further on alignment, by training the model to behave better when faced with complex tasks or tasks performed under pressure.
According to the company’s internal tests, Carbon makes roughly half as many unsupported claims as Argon. When elements are missing, it signals this in its deliverable or lists what remains to be provided. Faced with an ambiguous request, it searches for relevant sources and submits inconsistencies to the user before acting. Google also indicates that it is much less likely to try to circumvent rules to complete a task, without degrading the quality of the code produced. On the Gray Swan benchmark dedicated to indirect prompt injections, malicious instructions hidden in processed data, none of the tested attacks succeeded, whereas Argon already had a very low success rate.
Limited Access and No Public Timeline
Google is repeating the method followed during the presentation of Argon in early October. For now, Carbon is being offered to a portion of cyberdefense players, before being gradually opened to other trusted partners under the Fairwind program. This caution is partly due to the model’s profile: a tool capable of verifying that a flaw is exploitable can also be used to exploit it. Google is also participating in the voluntary pre-release model testing framework proposed by the U.S. government.
The company says it wants to gather feedback and adjust its safeguards before opening Carbon to developers, businesses, and the general public. No date is given, even though two weeks elapsed between the announcement of Argon and its public release.
Regarding pricing, Gemini 4 Carbon will be launched at the same introductory price as Argon, i.e., $2 per million input tokens and $10 per million output tokens, the token being the unit of text processed by the model. Cached input tokens receive a 95% discount. At the end of the launch period, the duration of which Google does not specify, the price will increase to $4 for input and $20 for output.
