Gemini 3 references found in code hint at Google DeepMind’s next AI model

References to a new model named Gemini Beta 3.0 Pro have been discovered in the Gemini CLI tool codebase, an open-source project maintained in part by Google employees. This suggests Google DeepMind is preparing a next-generation release following the current Gemini 2.5 Pro. While the 3.0 naming points toward a major update, the exact model name for public release is still uncertain.

What adds to the speculation is the previously spotted model named Kingfall, which delivered top-tier performance in early tests. Its real identity remains ambiguous—either an early version of Gemini 3 or a separate Gemini 2.5 Pro variant equipped with the long-rumored “Deep Think” feature. Deep Think, a capability teased in earlier leaks, is expected to offer more advanced reasoning and chain-of-thought abilities, particularly for Gemini on the web. This would align with Google’s ongoing strategy of expanding Gemini’s utility across platforms like Chrome and Workspace apps.

The leaked Gemini 3 reference matters particularly now, as it appears just ahead of major competitor announcements: xAI’s Grok 4, OpenAI’s possible GPT-5, and a rumored Claude update from Anthropic. The timing suggests Google might be positioning Gemini 3 as a direct counter to these launches, targeting users demanding stronger performance for coding, reasoning, and research use cases.

These findings were made through source analysis of the CLI tool, not through an official announcement. Until further confirmation emerges, Gemini 3 and Deep Think remain under internal testing, possibly both tied to Kingfall-level performance, which could indicate a shared architecture or parallel experimentation within DeepMind’s research teams.