Google Drops Gemini 3, Claims Model Beats GPT & Claude on Major Benchmarks

Google on Tuesday announced Gemini 3, its most advanced multimodal model to date, and a new agent-first developer platform called Google Antigravity.

The company says Gemini 3 raises the bar on reasoning, multimodal understanding and long-horizon planning, and will be rolled out across Search, the Gemini app, developer tools and enterprise products.

Alphabet CEO Sundar Pichai framed Gemini 3 as the next major step in the “Gemini era,” claiming rapid user adoption of prior generations and announcing broad availability of Gemini 3 Pro and a delayed, safety-screened Deep Think mode.

Google says the model will appear immediately in AI Mode in Search, the Gemini app, AI Studio, Vertex AI and other developer surfaces, and that Deep Think will be opened to safety testers before a wider release to Google AI Ultra subscribers.

Benchmark Performance

Google presented a battery of benchmark scores, asserting that Gemini 3 Pro outperforms Gemini 2.5 Pro across major tests, topping LMArena, advancing math and reasoning results, and showing strong multimodal performance on image and video tasks.

The company is also offering Gemini 3 Deep Think for higher-end reasoning and a separate “Deep Think” evaluation path that it says further improves scores on the most demanding benchmarks.

A central product reveal was Google Antigravity, an agentic development environment that ties Gemini 3’s reasoning and tool-use into an IDE-like surface where agents can plan, act and validate works autonomously across editor, terminal and browser. Google said Antigravity, plus Gemini CLI and AI Studio, will enable developers to build more autonomous, end-to-end agents and richer “vibe-coding” experiences.

Availability & Ecosystem

Google said Gemini 3 will ship at scale: in Search (AI Mode) for paying subscribers, across the Gemini app for general users, and via APIs for developers and enterprises in Vertex AI. The company also emphasised integration with third-party developer tooling and partner platforms, and signalled an ongoing cadence of additional Gemini 3 family releases.

Google stressed that Gemini 3 underwent an expanded safety evaluation, including independent assessments and engagement with external advisory groups. It highlighted improvements such as reduced sycophancy, stronger injection resistance and other hardening measures, and said Deep Think is being held for extra safety review before full public availability.

Rising Competition

The announcement positions Google to compete on three fronts simultaneously, packaged consumer experiences (Search and the Gemini app), enterprise AI (Vertex and Foundry) and developer/agent tooling (AI Studio, Antigravity).

If the claimed gains in reasoning, multimodal fidelity and long-context planning hold up in real-world workloads, Gemini 3 could accelerate adoption of agentic workflows and raise the bar for model-driven developer productivity, but execution, safety validation and enterprise uptake will determine commercial impact.