Gemini

• Gemini 3.8 Flashdefault

  gemini-3.8-flashReleased 2026-09Source

  Most capable Flash model, for agents and long coding tasks. Stable since 2026-09-02. thinking_level low, medium or high (default medium). Available on all Antigravity plans.

• Gemini 4 Argontop

  gemini-4-argonReleased 2026-10Source

  Google's most capable model. Announced 2026-09-30 with limited access; Gemini API GA 2026-10-13 (introductory pricing to 2026-11-30). 1M context; thinking default medium. Gemini app availability not confirmed.

• Gemini 3.5 Flash-Litebudget

  gemini-3.5-flash-liteReleased 2026-07Source

  Fastest, most cost-effective 3.5 model; thinking default minimal.

Where it runs

• Gemini app skills (invoke with /)
• Gems (personal accounts until November 2026; Workspace until March 2027)
• Google AI Studio and Gemini API
• Antigravity (desktop app and CLI)

Prompting notes

• Supply all context first and put the specific question or instruction at the very end; bridge with a phrase such as 'Based on the information above'. Source
• Separate prompt parts with XML-style tags or Markdown headings, and use one format consistently within a prompt. Source
• Gemini 3 and later answer directly and briefly by default: ask explicitly for a conversational or detailed reply. Source
• Set depth with thinking_level (low, medium, high; default medium on Gemini 4 Argon and 3.8 Flash) and pick the lowest level that meets the quality bar. Gemini 4 and 3.6+ ignore sampling parameters. Source
• Include a few varied examples with identical formatting; too many examples cause overfitting. Source
• In the Gemini app, save reusable instructions as a skill (type / and the name); Gems move to skills for personal accounts from November 2026. Source