Summary
Google's Gemini API: multimodal from the ground up (text, image, video, audio in and out), massive context windows (1M-2M tokens), and integration with Google's search grounding. The provider to pick when the task is multimodal or context-heavy.
Objectives
- 01State the Gemini model tiers.
- 02Explain what long-context specifically enables.
- 03Recognize grounding-with-search as a Gemini-native feature.
The Lesson
The tiers
Gemini 2.0 Flash: fast, cheap, 1M context. Gemini 2.0 Pro: strong general-purpose. Gemini 2.5 (as of 2026): frontier reasoning, extended context handling. Multimodal input works for images, video (as frame sequences), and audio.
Long context
1M tokens fits ~750K English words — every internal document of most companies. Enables patterns other providers cannot: whole-book summarization, whole-repo code understanding, whole-year of transcripts. Effective context degrades before 1M; measure on your task.
Key Ideas
- Gemini's advantage is multimodal + long context.
- 1M nominal ≠ 1M effective; measure.