Google launches Gemini 4 Carbon, its strongest model yet, and brings it to AI Mode in Search

Google has launched Gemini 4 Carbon for developers, businesses, scientists and consumers, and the company calls it its strongest model. The announcement came on November 10, 2026. Beyond the API, the model is starting to reach AI Mode in Search and the Gemini app for Google AI Ultra subscribers.

In the announcement on Google The Keyword, published November 10, the company says Carbon was built for tasks that push the limits of a model's intelligence. These include complex multistep coding, debugging and open-ended scientific research. To handle them, it spends more tokens than Gemini 4 Argon: it explores the problem more carefully, reasons more deeply and runs through long verification cycles.

Argon, released earlier, is still what Google recommends for most software engineering and knowledge work. Two weeks before the general launch, Carbon had already been made available to a group of cyber defense professionals through the Fairwind program.

Where the model is available and what it costs

Carbon is available through the API in Google AI Studio and in Gemini for business for all paying developers and businesses. Introductory pricing matches Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens.

Google AI Ultra subscribers will get the model gradually in Google Antigravity, the Gemini app and AI Mode in Search. In Antigravity, Google also opened a preview of the /swarm feature, which turns a single prompt into a team of coordinated agents that you can follow in real time.

Performance and document analysis

According to Google, Carbon took first place on CursorBench 4.0, on Perplexity's WANDR and on Zapier's AutomationBench. On those tests, its cost per task was more than 2 times lower than that of the next best external model. On the new Multimodal Knowledge Agent (MKA) benchmark, created with Artificial Analysis, Carbon came in second, with an average cost per task more than 6 times lower than the top model.

The company also says the model combines visual understanding, reasoning and exhaustive search to draw grounded conclusions from large sets of documents and dense data. It also follows strict schemas and formats when it generates spreadsheets, presentations and reports.

On safety, Google reports that the success rate of prompt injection attacks against Carbon was 0% on Gray Swan's IPI benchmark. The company acknowledges that resistance to this type of attack still needs to improve.

What this means for brand teams

Carbon is starting to power answers in AI Mode and the Gemini app for Google AI Ultra subscribers. The announcement does not describe any change in how those answers cite sources or surface brands. What Google highlights is the model's ability to search deeply and organize information from dense documents, which is worth watching in the answers where your brand shows up.