Google launches Gemini 3 with new coding app and record benchmark scores (techcrunch.com)

Google today launched Gemini 3, its most capable foundation model to date, available now in the Gemini app and AI search interface. Arriving seven months after Gemini 2.5 and amid rapid releases from rivals (OpenAI’s GPT‑5.1, Anthropic’s Sonnet 4.5), Gemini 3 claims a significant jump in reasoning and nuance. A research‑intensive variant, Gemini 3 Deepthink, will follow for Google AI Ultra subscribers after additional safety testing. On benchmarks the model set new records: 37.4 on the Humanity’s Last Exam (vs. GPT‑5 Pro’s 31.64) and it topped LMArena’s human‑led user‑satisfaction leaderboard. Google notes the Gemini app reaches 650 million monthly active users and has been used by 13 million developers, underscoring broad deployment. Alongside the base model Google released Antigravity, a Gemini‑powered, multi‑pane agentic coding interface that combines a chat prompt, command‑line, and a browser window to show live effects of code edits—an IDE-style workflow akin to Warp or Cursor 2.0. The combined offering signals two concrete implications: (1) frontier LLMs are making measurable leaps in multi-step reasoning and user satisfaction, and (2) agentic developer tooling is maturing into integrated, interactive environments that can orchestrate editors, terminals, and browsers. The release accelerates competitive pressure across the frontier LLM landscape while highlighting safety testing and deployment tradeoffs for higher‑capability “Deepthink” models.