Google releases Gemini 4 Carbon with enhanced coding and security features

Google has officially released Gemini 4 Carbon to paying developers and enterprise customers, marking the second iteration of its ultra-large model series. The company positions Carbon for complex tasks such as multi-stage code debugging and open scientific questions, while retaining Gemini 4 Argon for general programming. In the CursorBench benchmark, Carbon ranks first and operates at more than double the cost-efficiency of the next best model. However, in a separate bug hunt test, Carbon identified 57 of 105 bugs, trailing Fable 5.2’s 58, but at a cost of $31 compared to nearly $140 for the rival model.

Pushmit Kohli, VP of DeepMind, highlighted Carbon’s ability to perform formal verification of Signal’s security properties by analyzing machine code, a task completed in one month that typically takes teams decades. Google also noted that Carbon found two zero-day vulnerabilities at major software providers in under two hours via a Wizz program. In prompt injection tests, no attacks succeeded, though experts note the test may be exhausted. The model consumes more tokens due to extended reasoning and self-verification, resulting in higher absolute costs despite identical per-token pricing.

Google simultaneously introduced 'Swarm' in Antigravity, a feature that spawns a team of AI agents from a single prompt, with Carbon handling coordination and smaller models working in parallel. While Swarm offers speed advantages, industry leaders like OpenAI and Anthropic remain skeptical about the complexity and cost of coordinating agent swarms. Ethan Mollick noted that Carbon improves on Argon and offers a distinct AI perspective compared to Claude and GPT, describing it as a highly capable model with its own strengths and weaknesses.