Google launches Gemini 3 AI model

Google has launched Gemini 3, the latest version of its generative AI model, across several products including Search, the Gemini app, AI Studio, Vertex AI, and a new developer platform called Google Antigravity.

The company said Gemini 3 outperforms earlier versions in benchmarks for reasoning, mathematics, and multimodal tasks, with top scores on tests like LMArena and MathArena Apex.

Gemini 3 Deep Think, an enhanced reasoning mode, is being tested before wider release.

Google said the model can process text, images, video, audio, and code, and supports a 1 million-token context window.

Developers can access Gemini 3 through Google’s platforms and third-party tools such as GitHub and Replit.

The company did not disclose commercial launch dates for all features.

🔗 Source: Google

🧠 Food for thought

Implications, context, and why it matters.

Gemini 3 sits in the mid range

Gemini 3 Pro costs $2 per million input tokens and $12 per million output tokens for prompts of 200K tokens or less 1, placing it above Gemini 2.5 Flash‑Lite that Google calls its cheapest model 2.
It offers a 1 million token context window 3. If an input exceeds 200K tokens, billing switches to long‑context rates for all tokens 4, so context heavy work will cost more than the baseline.
Antigravity’s public preview mentions generous rate limits 5. The announcement shares no quotas, Service Level Agreements (SLAs), or general availability timelines 5, which makes production planning and cost comparisons hard.

Where agent ops plug into Antigravity

Agent ops vendors build tools to operate, monitor, and secure AI agents in production. They can tap Antigravity artifacts such as task lists and implementation plans with browser recordings to build integrations for observability (visibility into agent behavior and performance), audit trails, and compliance logging 5.
Antigravity runs agents asynchronously across editor, terminal, and browser surfaces 5. The announcement does not spell out guardrails (policy and safety constraints), security sandboxing (isolated execution environments), or evaluation frameworks (systematic testing and scoring) 5. Security startups can add behavior monitoring, permission controls, and runtime checks for enterprise use.
Antigravity works with multiple model providers, including Claude Sonnet 4.5 and GPT‑OSS 5. The announcement skips documented APIs for custom model routing or cost control 5. Agent ops platforms can pick models by task complexity, latency, and budget across providers.