Gemini 4 Carbon is aimed at tasks where a model has to carry out several operations in sequence: examine code, look for a vulnerability, produce a fix and verify it, or summarise lengthy documents. Google is making it available to more users, but in different ways depending on whether they use the Gemini app, its development tools or an offering for researchers. The question, then, is not just what Carbon can do, but under what conditions it can be entrusted with the work.
A model for lengthy tasks, not every use
Google describes Carbon as its most powerful model for complex, multi-step tasks, particularly software development, scientific research and some cybersecurity work. The company explicitly distinguishes these tasks from everyday uses: it says Gemini 4 Argon remains suitable for most knowledge and programming tasks. Carbon is designed to devote more computing power to a request, exploring several approaches and repeating verification steps.
This distinction matters in an agentic system. An agent does not merely write a response: it can use tools, examine files or carry out a series of actions to achieve a goal. The result then depends as much on the quality of the model as on its access to tools, the instructions it receives, the data provided and the oversight of its actions. A lengthy task also creates more opportunities to go down the wrong path.
Google highlights performance on CursorBench 4.0, WANDR, Perplexity’s test and Zapier’s AutomationBench. The company says Carbon tops these evaluations and costs, per task, more than twice as little as the next-best external model. These are results reported by Google for tests associated with specific uses: on their own, they do not show that an agent will achieve the same advantage in a team’s day-to-day work.
Fluency is no substitute for oversight
Figures reported by Frandroid from Artificial Analysis evaluations shed light on another dimension: the ability to give the right answer and the risk of making one up are not the same measure. On AA-Omniscience, Carbon gives the right answer to 58% of questions, according to these results, compared with 48% for Argon. But among questions the model answers incorrectly, the share of answers it makes up rather than admitting uncertainty rises from 4% for Argon to 11% for Carbon.
These rates apply to this particular test, not to all uses of Carbon. They nevertheless point to an operational precaution: an answer that appears precise is not the same as a verified one. For a legal document, a financial analysis or a code change, users need to be able to check the sources, calculations or changes made. The more complex the task, the more important it becomes to track each step.
Google has also announced a new multimodal test, the Multimodal Knowledge Agent, developed with Artificial Analysis to evaluate agents working on documents that combine text, tables, images and charts. Carbon comes second, according to Google, at an average cost per task more than six times lower than that of the top-ranked model. Here too, the result relates to a specific evaluation; it is no guarantee of accuracy for every business case.
Access depends on the use
For the general public, Google says Carbon is being rolled out in the Gemini app and in AI Mode in Search to Google AI Ultra subscribers. Pro subscribers therefore do not have the same access in these products. Developers and businesses can use Carbon through the Gemini API on paid plans, as well as through some of Google’s professional tools. Access through a consumer product and access via API are not interchangeable: the latter makes it possible to integrate the model into software or a process, but requires that integration to be built and maintained.
Google is also offering developers a preview of /swarm in Antigravity. A request can be split between several agents working in parallel, with different roles, such as designing an architecture, coordinating or auditing. This approach can speed up work by dividing it up, but it does not remove the need to check the result. Agents may share the same mistaken assumption or produce conflicting changes; supervision and testing therefore remain central to the process.
Researchers are another group with a distinct form of access. Google says it is extending to Carbon a programme that gives selected scientists free access to models for six months via the highest usage tier of Google Antigravity, subject to eligibility, capacity and selection. This is not general availability: access to the programme is granted as applications are considered.
Cyber capabilities need safeguards
Cybersecurity is one of the areas where the gap between capability and conditions of use is particularly significant. Google says Carbon can help identify vulnerabilities, understand how they interact and produce fixes. The company reports several results from its own evaluations and collaborations with security teams. It specifies that access to the version intended for this work is through its Fairwind programme, which is reserved for cybersecurity defenders.
Google also describes safeguards against malicious use and hidden instructions in content examined by an agent, a form of attack known as indirect prompt injection. The company reports good results on several security tests, while acknowledging that resistance to injections still needs to improve against certain sophisticated attacks. This limitation matters to any organisation that allows an agent to read web pages, emails or files: a hostile instruction can be hidden in the data it consults without being stated directly by the user.
Carbon therefore expands the range of tasks Google wants to automate, particularly those involving several steps and the use of tools. But access varies by product, and published performance figures describe tests or evaluations as reported by their authors. Before deciding to entrust it with a task, the whole process needs to be considered: accessible data, permitted actions, cost, available logs and the possibility of human oversight. The ability to produce an answer or a fix is only part of the work.
References
- Google, “Gemini 4 Carbon is now available”
- Frandroid, “You can now use Gemini 4 Carbon (if you’re willing to pay a lot)”
Google offices in Sunnyvale, California. Illustrative photo.
