Google has launched Gemini 4 Carbon, an AI model engineered for complex, multi-step reasoning tasks that delivers top performance across three major agent benchmarks while reducing per-task costs to less than half of competing models, according to an announcement by Koray Kavukcuoglu, Google DeepMind Senior Vice President. The release targets applications including complex coding, open scientific research, and cybersecurity.

Performance Across Benchmarks and Early Applications

Gemini 4 Carbon ranked #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench. The model is designed to perform additional verification steps before answering, which results in higher token usage compared to its sibling model Gemini 4 Argon. Google recommends Argon for most software engineering and knowledge work, reserving Carbon for tasks that test the limits of model intelligence.

In life sciences applications, Google researchers reported finding, on average, 40 times more remote homologs than before across seven protein families with low homology. This data is used for protein structure prediction models such as AlphaFold 3.

The model's security capabilities showed early promise when the Wiz team used Carbon to identify zero-day vulnerabilities at two major SaaS providers within two hours.

Google built a multimodal knowledge agent benchmark with Artificial Analysis to handle dense documents such as charts, blueprints, and tables. Carbon ranked #2 in this evaluation, though Google emphasized that Carbon's average per-task cost was more than six times lower than the #1 model.

In finance-related evaluations, Carbon matched the top model in Hevia's end-to-end financial agent evaluation and ranked #1 among 23 frontier model configurations in Valyasnyi's general financial intelligence benchmark.

Pricing and Access Structure

Initial API pricing for Carbon and Argon is set at $2 per 1 million input tokens and $10 per 1 million output tokens through November 30. After the introductory period ends, Carbon's input fee doubles to $4 and its output fee doubles to $20.

Consumer access is available through a $100 monthly Google AI Ultra subscription, which provides Carbon access in the Gemini app, Search's AI Mode, and Antigravity platform. The Antigravity platform will offer a preview of a '/swarm' feature that forms multiple agent teams with a single prompt. Developers can assign Carbon to tasks requiring advanced reasoning, such as architecture design or auditing, while delegating parallel tasks to lighter models.

Research Program Expansion

Google is offering 10,000 scientists six months of free access to both Argon and Carbon through the Antigravity platform. The company has been accepting applications for the past four weeks, with thousands already enrolled. Current and future research program participants will receive Carbon access alongside previously available models, with Google providing Argon for general tasks and Carbon for high-difficulty reasoning to broaden its developer and researcher base.

Carbon's combination of benchmark performance, cost reduction compared to competitors, and early demonstrations in protein analysis and security applications positions it as a general-purpose AI model with particular utility for complex reasoning tasks across scientific and technical domains.

Sources

  1. Google unveils Gemini4 Carbon, boosts performance and cuts costs — DigitalToday