Google Gemini 4 Carbon takes the lead on AI intelligence index

Reclaiming the frontier

Google is staging a comeback in the high-end artificial intelligence market with the release of Gemini 4 Carbon. As reported by AI Magazine, the model has secured the top spot on the Artificial Analysis Intelligence Index v5. This puts it ahead of major competitors like Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra.

The shift marks a significant moment for Google, which saw its early lead in the field overtaken by OpenAI and Anthropic in recent years. Under the leadership of Koray Kavukcuoglu at Google DeepMind, the company is now pushing back into the frontier of what these models can achieve. The goal is to move beyond simple chat interactions toward more difficult, long-horizon engineering problems.

The model is designed to handle sustained agentic workflows, which are systems capable of performing multi-step tasks autonomously. Kavukcuoglu noted that internal teams have observed Carbon operating for several days to navigate obstacles and find solutions. It is specifically built to manage intricate engineering work, such as identifying the source of errors and fixing minor defects within massive software repositories and low-level systems.

Performance and benchmarks

Gemini 4 Carbon’s score of 59 on the Intelligence Index v5 is driven by its performance in complex reasoning, knowledge, and scientific tasks. While it is not the leader in every single category, it shows large gains in coding and knowledge work compared to previous Gemini models. It also leads the AA-Omniscience v1.1 benchmark while maintaining a relatively low 11% hallucination rate.

However, the model does have some specific limitations compared to other specialized tools. For example, it trails SpaceXAI’s Grok 4.7 and Meta’s Muse Wonder on the AA-Briefcase knowledge-work benchmark. Artificial Analysis also noted that while the model is faster than average, it can be very verbose, meaning it may provide more text than is strictly necessary for a given prompt.

Despite these nuances, the model shows significant improvements in security and reliability. Google reported that Carbon achieved more than a 2x uplift in success rates for generating secure code over the previous Gemini 4 Argon. Additionally, the Chrome team noted that Carbon was over ten times more effective than a large frontier model from a different company when creating high-quality software fixes.

Technical specs and costs

The model is built to be versatile, supporting multimodal inputs such as text, image, speech, and video. It produces text outputs and features a 1 million token context window, which allows it to process large amounts of information at once. This capability is essential for the long-horizon engineering tasks the model was designed to solve.

From a pricing perspective, the model presents a mixed profile. Artificial Analysis describes it as somewhat expensive when compared to other models with similar capabilities. However, it does offer some cost advantages over specific competitors, with standard task costs running 23% cheaper than Claude Opus 5.5 at a price of US$7.76.

To encourage adoption, Google is offering 50% launch discounts for the model. These incentives are part of a broader strategy to integrate Gemini into various enterprise and consumer ecosystems. The model is currently being rolled out to selected users and is not yet available to the general public.

Partnerships and availability

Google is leveraging its existing partnerships to bring Gemini 4 Carbon to a wider audience. The company works with Apple to provide the models and infrastructure needed to power the next generation of Siri and Apple Intelligence. It also maintains deep integration with Samsung across the Galaxy AI consumer hardware ecosystem.

In the enterprise space, Google has an expanded alliance with Accenture. This includes the Gemini Enterprise Acceleration Program, which aims to deploy agentic AI workflows across operations for Fortune 500 clients. These partnerships highlight how the model's ability to handle complex, multi-step tasks is being prioritized for business use cases.

For the general public, the rollout will happen in stages. Alphabet and Google CEO Sundar Pichai stated that the model will be expanded to more partners through the Fairwind Program before reaching wider availability. For now, the model serves as a demonstration of the pace at which Google is pushing the current intelligence frontier.