Last updated:
Gemini
Google DeepMind's natively multimodal model family, notable for very long context windows and strong reasoning benchmarks.
Model Facts
Latest model: Gemini 4 Argon
- Provider
- Google DeepMind
- Released
- Sep 2, 2026
- Context window
- 1M tokens
- Weights
- Closed
Benchmarks
95.3%
GPQA Diamond
3.8 Flash
89.2%
ARC-AGI-2
3.8 Flash
Full Model List
| Model | Released | Note |
|---|---|---|
| Gemini 4 Argon | Sep 30, 2026 | limited preview for vetted cyber defenders only |
| Gemini 3.8 Flash | Sep 2, 2026 | latest stable release |
| Gemini 3.5 Flash | May 19, 2026 | announced at Google I/O; the only 3.5-family model to ship so far |
| Gemini 2.5 Pro | Mar 25, 2025 | topped LMArena at launch |
| Gemini 2.0 | Dec 11, 2024 | agentic-focused generation |
| Gemini 1.5 Pro | Feb 15, 2024 | introduced the 1M-token context window |
| Gemini 1.0 | Dec 6, 2023 | first Gemini generation |
| Gemini 3.1 Pro | – | Pro-tier model: 94.3% GPQA Diamond, 77.1% ARC-AGI-2 |
- Official model page:
- deepmind.google/technologies/gemini
- Official documentation:
- ai.google.dev/gemini-api/docs
- Other sources:
- Gemini (Wikipedia)
- Gemini 3.8 Flash on OpenRouter
What it is
Gemini is Google DeepMind's natively multimodal model family. The latest stable release is Gemini 3.8 Flash (September 2, 2026); the most capable Pro-tier model that is generally available is Gemini 3.1 Pro. Gemini 3.5 Pro was announced at Google I/O in May 2026 but had still not shipped as of the latest reports, and Gemini 4 Argon (September 30, 2026) is limited to vetted cyber defenders for now.
Context and limits
Gemini 3.x models take up to about 1M input tokens; Gemini 3 Pro-class models produce up to 64K output tokens, and Gemini 4 Argon raises the output limit to 1M tokens.
Where these numbers come from
The headline figures are for Gemini 3.8 Flash — 95.3% on GPQA Diamond and 89.2% on ARC-AGI-2 — as reported by third-party trackers. For comparison, Gemini 3.1 Pro scores 94.3% and 77.1% on the same tests. Treat vendor and aggregator numbers as indicative, since harnesses differ.
Notable traits
- Very long context windows (1M tokens) across the family.
- Natively multimodal: text, images, audio and video.
- Closed weights — available through the Gemini app and API, with open "Gemma" models as a separate line.