Google’s new Gemini 4 Carbon landed yesterday, and the benchmark numbers are worth paying attention to. It ranks first on three major independent evaluations of real-world AI workflow performance, including CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench. It achieves those rankings at more than twice the cost efficiency of the next best competing model. That combination is unusual.
The price has not changed. Carbon costs the same as Gemini 4 Argon, Google’s previous flagship model released just weeks ago. Developers and enterprises pay $2 per million input tokens and $10 per million output tokens. For context, a “token” is roughly three-quarters of a word. Top benchmark performance, no price increase.
Carbon is built for work that pushes a model hard. Legal analysis of large document sets, financial research across complex datasets, multi-package software debugging, and open-ended scientific discovery are the primary use cases. This is not a general chatbot upgrade. It is a model designed for extended, multi-step reasoning tasks that require sustained depth and accuracy across many sequential decisions.
The safety results deserve attention. Carbon achieved a 0% prompt injection attack success rate on an independent security benchmark, the first time any frontier model has hit that mark. Prompt injection is an attack method where malicious instructions are hidden inside content the AI processes, with the goal of hijacking what the AI does next. For businesses running AI agents inside real workflows, that number matters a great deal.
A preview feature called /swarm changes what is possible for large projects. Available inside Google’s Antigravity platform, Swarm converts a single prompt into a coordinated team of AI agents working in parallel, with separate agents reviewing and testing each other’s output. Think of it as assigning a complex task to a team rather than one person. For large software migrations or multi-document research tasks, that parallelism can compress hours of sequential work into a fraction of the time.
Carbon is available now through Google AI Studio and Gemini for Business. Google AI Ultra subscribers get access through the Gemini app and AI Mode in Search.
Three things stand out from a business perspective. First, the cost-to-performance ratio makes the economics of AI automation harder to ignore. If your team is spending hours on tasks Carbon handles accurately for cents, the math will not hold up much longer. Second, the security improvements make Carbon a more credible option for regulated industries like healthcare, finance, and legal services. Third, the multi-agent swarm capability signals where this is all heading. Not one assistant answering questions, but coordinated AI teams executing complex, multi-step work on their own.
The window to get ahead of this shift is still open. It is narrowing.
Want to explore how AI agents could benefit your business? Let’s talk.