Google Ships Gemini 3 Across Search, Cloud and a New Agentic IDE (Nov 2025)

On 18 November 2025 Google released Gemini 3 Pro, the first model of its third-generation family, in a post co-signed by Sundar Pichai, Demis Hassabis and Koray Kavukcuoglu. Unlike previous generations, which reached Google's consumer surfaces months after their API debut, Gemini 3 arrived on day one in the Gemini app, in AI Mode in Search for paying subscribers, in Google AI Studio and Vertex AI for developers and enterprises, and inside third-party tools such as Cursor, GitHub, JetBrains and Replit.

Google's published numbers put the model at the top of several public leaderboards: 1501 Elo on LMArena, 37.5% on Humanity's Last Exam without tools, 91.9% on GPQA Diamond, 76.2% on SWE-bench Verified and 81% on MMMU-Pro. A separate Gemini 3 Deep Think reasoning mode, slated for Ultra subscribers "in the coming weeks", scored 41.0% on Humanity's Last Exam and 45.1% on ARC-AGI-2 with code execution. Google also described the release as its most secure model to date, citing reduced sycophancy, stronger resistance to prompt injection, and pre-release evaluations run with the UK AI Security Institute and independent assessors.

The most consequential part of the launch for enterprise buyers may be Google Antigravity, a new agent-first development environment in which Gemini 3 plans and executes multi-step coding tasks with access to an editor, a terminal and a browser. It is Google's most direct answer to Claude Code, GitHub Copilot's agent mode and Cursor, and it signals that the company sees autonomous software agents, not chat, as the primary commercial vehicle for its frontier model.

The timing matters. Gemini 3 landed six days before Anthropic's Claude Opus 4.5 and three weeks before OpenAI responded with GPT-5.2, compressing what used to be a yearly cadence into a single quarter. Google's decision to push the model into Search immediately also indicates growing confidence in factuality controls, since Search errors are visible to billions of users. Within days, however, security researchers published a working data-exfiltration chain against Antigravity using an indirect prompt injection hidden in web documentation, a reminder that the gap between model-level robustness claims and product-level safety remains wide.

For organisations evaluating frontier models, the practical picture is now one of three broadly comparable vendors leapfrogging each other on public benchmarks every few weeks. The differentiators that will hold up over time are not benchmark points but governance features: evaluation transparency, enterprise data controls, agent permission models and the maturity of the vendor's incident response.

What it means for leaders

• Treat model selection as a portfolio decision, not a bet. With Google, Anthropic and OpenAI trading the lead quarterly, architectures should keep model routing abstracted so that a switch is a configuration change rather than a re-platforming.
• Separate model risk from product risk. Google's prompt-injection improvements apply to the model; Antigravity's terminal and browser access create new attack surface at the product layer. Map both in your NIST AI RMF risk register.
• Agentic IDEs need explicit policy before adoption. Define which repositories, secrets and network destinations an agent may touch, require human approval for command execution, and align controls with OWASP LLM Top 10 items on prompt injection and excessive agency.
• Ask for the evaluation evidence. Google references third-party assessments; procurement teams should request the scope and results as part of due diligence, consistent with ISO/IEC 42001 supplier-control expectations.
• Watch Search integration effects. AI Mode answers built on Gemini 3 will shape how customers find your organisation; digital and communications teams should monitor how their content is being summarised.