Google launches Gemini 4 Carbon for the hardest tasks

What's new

Google has made Gemini 4 Carbon available to developers, businesses, researchers and everyday users. Google calls it its most capable model and points to cybersecurity and long agentic tasks, where the AI works independently across many steps, such as knowledge work, software development and scientific reasoning.

Carbon follows Gemini 4 Argon, which Google still recommends for most coding and knowledge work. Carbon is for the toughest tasks: complex multistep work in coding and debugging, as well as open-ended research.

Paying developers and businesses get it in Google AI Studio and Gemini for business. Google AI Ultra subscribers get it in Antigravity, the Gemini app and AI Mode in Google Search. The price of the Ultra subscription has not been disclosed, and the sources do not make clear whether Carbon is available in Denmark and the EU.

What's smart

Carbon uses more tokens (chunks of text) than Argon to explore a task thoroughly, reason deeply and run extended verification loops, meaning it checks its own work several times.

Google is also previewing /swarm in Antigravity. The feature turns a single prompt into a team of agents that you can follow in real time. Carbon can take on roles such as architect, coordinator and reviewer, while faster, lighter models work in parallel. Google's examples are an interactive 3D exhibit about the construction of the Eiffel Tower and a reconstruction with 906,000 triangles of Google's Platform 37 office, built from seven PDFs and 206 drawing views over 10 overnight loops.

Cyber defenders got access two weeks before launch through the Fairwind program, and cybersecurity professionals can still apply for access.

Cheaper or better?

The introductory price is $2 per million input tokens and $10 per million output tokens, the same as Argon. According to a summary of Google's footnote, it applies through November 30, 2026. After that, the price is $4 for input and $20 for output per million tokens. Because Carbon uses more tokens, the bill per task can end up higher than the token price alone suggests.

Google says Carbon ranks No. 1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at less than half the cost per task of the next-best external model. That model is not named. On the MKA benchmark, Carbon ranks No. 2 and is more than six times cheaper than the top model. On Balyasny Asset Management's finance benchmark, it placed first among 23 frontier model configurations. The figures are Google's own or its partners' choices and have not been independently verified.

Another picture comes from SemiAnalysis' GDPx test with 206 tasks, as reported by Let's Data Science. Here Carbon ranks No. 2 behind Claude Opus 5.5, with GPT-6 Astra in third place. The gap between first and second place is small. Carbon costs about $16 per task, compared with about $51 for Claude Opus 5.5 and about $54 for GPT-6 Astra. If cybersecurity, biology and alignment are removed, GPT-6 Astra moves to No. 1. According to the coverage, the report attributes this to safety guardrails in the Claude and GPT models.

What it's good for

Carbon is for heavy tasks where you want an AI to work long and thoroughly: difficult coding and debugging, research, cybersecurity and financial analysis. For everyday coding and knowledge work, Google still recommends Argon.

Google acknowledges that the model's resistance to prompt injection, where hidden instructions in a text try to trick the AI, could be improved. Broader scientific benchmarking is still under way, and researchers' access depends on suitability, capacity and selection.