Gemini 4 Carbon Beats GPT-6 Astra and Claude Opus 5.5 as the World’s Top AI Model

Google has unveiled Gemini 4 Carbon, a newly developed AI model that has outperformed both OpenAI's GPT-6 Astra and Anthropic’s Claude Opus 5.5 in a major intelligence ranking. The new model scored 59 on the Artificial Analysis Intelligence Index, surpassing both GPT-6 Astra (57) and Claude Opus 5.5 (58). While Carbon is currently not available to the public, Google plans to gradually release access to a select group of cybersecurity experts before expanding to developers, businesses, and consumers.

What sets Carbon apart is its exceptional performance in coding, AI agents, and cybersecurity tasks. The model demonstrates significant improvements in long-running coding tasks, AI agents, and cybersecurity defenses. Google claims that Carbon writes more secure code in its internal tests and shows fewer signs of misalignment during live coding tasks compared to its predecessors.

Carbon is designed for long-term, complex engineering problems, with the ability to run for days to complete intricate tasks such as tracing and fixing bugs in large codebases. The model recently outperformed its predecessor, Gemini 4 Argon, which achieved an Artificial Analysis Intelligence Index score of 50. Carbon's score of 59 marks a substantial leap forward, with notable gains in long-term context understanding and overall performance.

In terms of cybersecurity, Carbon is considered a formidable contender, leading in vulnerability discovery, exploitation checking, and patching. The model scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. Additionally, Google's Chrome team reported that Carbon performed 10 times better than another large frontier model at generating high-quality patches.

Despite its impressive performance, Carbon does face stiff competition from GPT-6 Astra, which leads in FrontierSWE 2 (65.5%), OSWorld-2.1 (87.2%), and the Vals Index v2.2 (70.4%). However, Carbon still manages to outperform GPT-6 Astra in several areas, including long context reasoning and security-related benchmarks.