AI library

Models

Gemini (Google)

Last reviewed 15 October 2026

Overview

Gemini is Google DeepMind's model family, used in the Gemini app, Google products, and through the Gemini API and Vertex AI. Gemini is natively multimodal (text, images, audio, video) and known for long context. Google also publishes the open-weight Gemma models (see open-weight models).

Key concepts

  • Tiers: Pro (most capable), Flash (fast workhorse), Flash-Lite (cheapest), plus specialised variants such as Flash Cyber.
  • Gemma: Google's open-weight family, separate from Gemini.
  • Current lineup (as of 2026-10-11): Gemini 3.8 Flash (GA 2026-09-02) is the newest generally available text model, aimed at long-horizon coding, agents, and enterprise workflows: 1M input and 64K output tokens, text, image, video, audio, and PDF input, thinking levels, and computer use (preview). Gemini 3.1 Pro Preview is the newest Pro. The July 2026 release added 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, with no 3.5 Pro. Gemini 3.7 Flash (released 2026-08-13) and 3.5 Flash (2026-05-19) were deprecated on 2026-10-08; their requests are now routed to 3.8 Flash and 3.6 Flash. Gemini 4 Argon (gemini-4-argon) is the most capable model: first released to cybersecurity partners in the Fairwind Program on 2026-09-30, then publicly on 2026-10-13 (paid API tier, Gemini app for Pro and Ultra, Antigravity). On the Gemini API it is a stable, GA model. It has a 1M-token limit shared by input and output; output defaults to 64K tokens per synchronous request, up to 1M with "long decode". Thinking levels are low, medium, and high. It supports implicit caching only (no explicit caching), and computer use is still a preview feature.
  • Gemini 4 Argon safety: Google applied its strongest Frontier Safety Framework safeguards for cyber and biology (input screening, activation probes, multi-step monitoring), and reports 0.7% attack success on Gray Swan's indirect prompt injection benchmark. See security.
  • Robotics: Google pitches Argon itself as a physical reasoning model for robotics (87% on ERQA), not a separate robotics model.
  • Voice and speech: Gemini 3.8 Live (and a Live Extended Thinking variant) from 2026-09-15; 3.8 Flash TTS and Flash-Lite TTS GA on 2026-09-22.
  • Image models ("Nano Banana"): Flash-class image generation and editing models. The current one is gemini-nano-banana-2.1 (GA 2026-10-06): panoramic ratios up to 8:1, 1K–4K output at every ratio, up to 14 reference images.
  • Antigravity: Google's agent preview (antigravity-preview-09-2026 since 2026-09-17) uses PascalCase tool parameters and line-range file edits.

How to use it

  • Flash models are the usual starting point for high-volume work.
  • Watch the deprecations page: deprecated Flash models may be routed to a newer model automatically, so pinning a model name does not guarantee the same model.
  • Use Gemini for heavy video and audio input, where it has long been strong.
  • Prices per 1M tokens (input / output, paid tier, as of 2026-10-07; Argon as of 2026-10-14): 4 Argon $2 / $10 until 2026-11-30, then $4 / $20; 3.8 Flash $0.75 / $3.75 (doubling in 2027), 3.5 Flash $1.50 / $9.00 (deprecated 2026-10-08), 3.1 Pro Preview $2 / $12 ($4 / $18 above 200K), 2.5 Pro $1.25 / $10, 2.5 Flash-Lite $0.10 / $0.40. See cost and caching.

Current state

  • 2026-10-15: The Gemini API model page labels gemini-4-argon Stable, and the changelog lists it as GA from 2026-10-13. No explicit caching; long decode up to 1M output tokens.

  • 2026-10-14: Google DeepMind and Singapore's government announced an on-device scam detector built on a fine-tuned Gemma model (not yet released). See open-weight models.

  • 2026-10-13: Gemini 4 Argon became publicly available on the paid Gemini API at introductory prices ($2 / $10, doubling after 2026-11-30). Google claims #1 on the Vals Index and 87% on ERQA. Companion posts cover its safety evaluations and robotics uses.

  • 2026-10-08: Gemini 3.7 Flash and 3.5 Flash deprecated, with requests routed automatically to 3.8 Flash and 3.6 Flash; no shutdown dates. The Deep Research agent deep-research-pro-preview-12-2025 shuts down on 2026-10-23 (replacements: deep-research-preview-04-2026 or deep-research-max-preview-04-2026).

  • 2026-10-07: The pricing page lists Gemini 3.1 Pro Preview as the newest Pro model, and Gemini 2.5 Flash-Lite as the only Flash-Lite model with a price. It also lists Gemini 3.7 Flash (released 2026-08-13), priced the same as 3.8 Flash.

  • 2026-10-06: Nano Banana 2.1 GA; gemini-3.1-flash-image is deprecated, with no shutdown date yet.

  • 2026-09-30: Gemini 4 Argon released only to cybersecurity partners.

  • 2026-09-18: Gemini 2.5 models restricted to existing active users; Google directs new projects to 3.5 Flash-Lite or 3.8 Flash.

  • 2026-09-02: Gemini 3.8 Flash GA, at introductory prices that double on 2027-01-01.

Open questions

  • Smaller Gemini 4 models (Flash, Flash-Lite) have not been announced (checked 2026-10-15). Answered on 2026-10-15: Argon itself is Stable and GA on the Gemini API. Its status on the Gemini Enterprise Agent Platform (Vertex AI) was not checked.
  • Google directs new projects to Gemini 3.5 Flash-Lite, but the pricing page (2026-10-07) lists no price for it or for 3.6 Flash.
  • Is traffic routed from deprecated 3.5 Flash to 3.6 Flash billed at the 3.6 Flash price? Neither the deprecations page nor the changelog says (checked 2026-10-12).
  • Nano Banana 2.1 per-image prices were not on its model page on 2026-10-07.

Sources