Gemini 4 Carbon Gives Google an Edge in AI Automation

Engineering and automation gains strengthen Google's competitive case, although advantages over OpenAI and Anthropic vary by task.

For people already using ChatGPT or Claude to write code or automate work, another AI model raises a practical question: how much more of the job can it actually finish?

Google announced Gemini 4 Carbon , targeting sustained engineering work, with initial access restricted to selected cyber defenders. It now leads Artificial Analysis' Intelligence Index v5, which combines coding, knowledge work, science and general knowledge.

As of October 29, Carbon scores 59 at high reasoning, ahead of 58 for Anthropic's Claude Opus 5.5 and 57 for OpenAI's GPT-6 Astra, both tested at maximum reasoning. The narrow lead becomes more useful when broken down by the work each model handles best.

Built for harder engineering work

Google says Carbon can spend days investigating and patching difficult bugs, using software tools to work through failed approaches.

Koray Kavukcuoglu, senior vice president of Google DeepMind and Google's chief AI architect, wrote in the announcement:

Designed to take on much harder, long-horizon engineering problems, Carbon is built for sustained agentic workflows.

Google's published scores include 88.2% on SEC-bench Pro for finding vulnerabilities and 54.8% on ExploitGym for validating exploits. On CWE-bench, which tests vulnerability repairs, Carbon scored 71%, against Astra's 68% and Opus 5.5's 67%. Together, these results point to potential across investigating security flaws and producing fixes.

Where the comparison becomes practical

On AutomationBench-AA, which measures progress through business workflows spanning multiple apps, Carbon scored 78%, compared with 70% for Opus 5.5 and 68% for Astra. For people connecting AI to their systems, that suggests potential to carry more work from retrieving information to updating records, with fewer manual handoffs.

Coding remains more competitive. In Artificial Analysis' coding-agent evaluation, Carbon with Google's Antigravity CLI scored 59, behind 62 for Opus 5.5 in Claude Code. Those scores reflect both the model and its tools.

Astra also led Carbon on GDP.pdf, a professional document-analysis test, scoring 31% against 28%. Carbon additionally offers a million-token context window and text, image, video and speech inputs, giving developers scope for workflows combining different media. The reason to choose it therefore depends on the task.

The cost of getting work done

Google's introductory API prices are $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period. Astra's standard short-context API rates are $10 and $50, respectively.

Yet lower token rates do not guarantee a lower bill. Using standard prices, Artificial Analysis measured $7.76 per Intelligence Index task for Carbon, against $10.04 for Opus 5.5 and about $6.25 for Astra. Astra used fewer output tokens. These benchmark costs show why usage matters alongside the advertised rates.

Source: Artificial Analysis, October 29, 2026. Standard prices; benchmark task costs.

Google has not announced a broader release date. Carbon's opportunity is to turn its engineering and automation gains into more work completed reliably at a competitive cost. That would give existing AI users a practical reason to add Gemini to their systems.