Start building with Gemini 3

Source: DeepMind Date: 2025-11-18 URL: https://deepmind.google/blog/start-building-with-gemini-3/

Summary

Google developer launch post for Gemini 3 Pro: 54.2% on Terminal-Bench 2.0 (tool use in terminal), ELO 1487 on WebDev Arena, new records on MMMU-Pro and Video MMMU, 1M context window. Surpasses Claude Sonnet 4.5 and GPT-5.1 per post’s claims. Pricing: $2/$12 per million input/output tokens (≤200K context). Introduced “vibe coding” — natural language to interactive applications. Available on AI Studio, Vertex AI, and Google Antigravity.

Implications

Terminal-Bench 2.0 at 54.2% is the agentic coding claim. Terminal tool use — running commands, navigating codebases, executing scripts — is what separates a code suggestion model from an autonomous coding agent. 54.2% on a benchmark explicitly testing this capability is a direct message to Devin, Cursor agent mode, and OpenAI Codex.

Named competitor comparisons (Claude Sonnet 4.5, GPT-5.1) in a product post are unusual. Google explicitly citing outperforming Claude and GPT in a launch post signals competitive intensity. The specific claim needs verification — model version timing and benchmark selection matter — but the naming signals Google is directly targeting Anthropic’s enterprise coding market.

“Vibe coding” framing is the consumer developer play. Positioning natural-language-to-application as “vibe coding” targets the same audience as Bolt, v0, and Lovable: non-expert developers who want to build without deep coding knowledge. That’s a large untapped market.

Watch:

• Independent Terminal-Bench 2.0 results — Google’s claims vs. Claude and GPT need third-party validation
• Google Antigravity as a platform: it’s mentioned here as a developer tool surface, which may explain the content-less page from the November announcement
• $2/$12 pricing vs. Claude Sonnet and GPT-4o at comparable capability tiers