AI News
AI News AgentModel releaseGoogle4 min read

Google launches Gemini 4 Carbon for complex tasks

Google is making Gemini 4 Carbon available to developers, businesses, scientists and AI Ultra subscribers. The model is designed for complex programming, cybersecurity, document analysis and scientific research, with a starting price equal to that of Gemini 4 Argon.

Google now lets you use Gemini 4 Carbon, an AI model designed for work that requires more reasoning, more steps and more verification than typical tasks. It is available to developers, businesses, scientists and Google AI Ultra users.

Carbon follows Gemini 4 Argon. Google presents Argon as the right choice for most programming and knowledge work, while Carbon targets more demanding projects: complex debugging, open-ended scientific research, analysis of dense documents and workflows in which multiple agents collaborate over long periods.

The practical difference is that Carbon can dedicate more tokens, meaning more processing space, to exploring a solution, checking it and correcting errors before delivering it.

A model for long and complex tasks

According to Google, Carbon ranks first in three independent tests focused on real-world tasks: CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench. The company says it does so at a per-task cost more than two times lower than the next-highest-ranked external model.

It also ranked second in the new Multimodal Knowledge Agent benchmark, created with Artificial Analysis to measure how models work with documents that combine text, graphics, images, diagrams and tables. In this test, Google says Carbon costs more than six times less per task on average than the leading model.

In practice, this could help you review contracts, summarize financial information, analyze spreadsheets or prepare reports that follow a specific format. It does not just answer questions: it can retrieve information, organize it and generate documents, presentations or tables with a defined structure.

Programming and agent teams

Carbon is designed for tasks such as migrating several software packages, investigating the source of a failure affecting different systems or redesigning an application's architecture. For building websites, Google says the model can write code, run it, check the visual result and roll back when it detects errors.

Google is also previewing /swarm in Antigravity. The feature turns an instruction into a coordinated team of agents that you can follow in real time. Carbon can handle decisions that require more reasoning, such as designing the architecture, coordinating the work or auditing changes, while faster models work in parallel.

That changes the approach: instead of asking AI to write an isolated function, you can assign it an entire project and review how it divides up the tasks and verifies the results.

Cybersecurity capabilities

In cybersecurity, Google says Carbon is its most capable model so far for detecting vulnerabilities, studying how they can be chained across systems and generating patches that verify the root cause of the problem.

As part of the Fairwind program, the company reports strong results in tests covering the detection, exploitation and remediation of vulnerabilities. In Project Fortify, its internal analysis and repair system, Carbon achieved 100% coverage in scanning and 95% accuracy in classification, according to Google's benchmark.

The company also says Carbon found 0-day vulnerabilities in the infrastructure of two major software-as-a-service providers in under two hours during a Wiz evaluation. Google has not yet identified the flaws because the responsible disclosure process is still open.

These capabilities have dual uses. They can help protect software, but they can also increase risk if used to attack systems. Google says its defenses blocked 99% of attack paths in simulated tests and reduced the intrusion rate per attempt to 0.8%, compared with 3.1% for Argon. These are internal results or findings from specific external evaluations, not a guarantee of absolute security.

Scientific research and mathematics

Carbon is also designed to work with experimental data, test hypotheses and solve problems in physics, mathematics and biology. Google says Axiom Bio considers it the best-performing model it has for applied scientific research and that it found toxicity mechanisms the company had not located before.

In another project, Google researchers combined Carbon with formal verification agents to prove security properties of code used by Signal's cryptographic protocol. The system generated millions of lines of code in Lean, a language used to write mathematical proofs, and completed in weeks a task that, according to Google, could historically require years of work from specialized teams.

To predict protein structures, Carbon located remote homologs in large sequence archives and structural databases. Across seven protein families with few known matches, Google reports an average 40-fold increase in the homologs retrieved. The larger-scale evaluation is still underway.

Pricing and availability

Gemini 4 Carbon is now available to developers and businesses through the API in Google AI Studio and Gemini for business. Its starting price is the same as Argon's: $2 per million input tokens and $10 per million output tokens.

It is also rolling out gradually to Google AI Ultra subscribers in:

  • Google Antigravity
  • The Gemini app
  • AI Mode in Search

Scientists can request six months of free access to Carbon and Argon through Google's Antigravity research program, subject to requirements, capacity and selection. Cybersecurity teams can request access to Carbon for defense through Fairwind.

For you, the main difference is not having another AI for answering simple questions. Carbon is designed for work where planning, review and the ability to maintain a task across many steps matter. What remains to be seen is whether its lab results hold up in everyday projects, how much it really costs to run long workflows and how its defenses evolve as agents gain more autonomy.