#3Google DeepMind

EmbeddingGemma 2

An open-weights embedding model from Google DeepMind that maps text, code, images, audio and video into one shared vector space. It is built on the Gemma 4 architecture and made to run on-device. Google calls it the most capable model for on-device multimodal embeddings, built from the same technology as its Gemini Embedding models.

• 740M parameters total; text-only use needs as little as 270M
• Text, code, image, audio and video in one embedding space
• 8K-token context window, 4x EmbeddingGemma 1
• 768-dim vectors, truncatable to 512, 256 or 128 (Matryoshka)
• MTEB Code 78.68, up from 68.76 (Google-reported)

AccessOpen weights on Hugging Face and Kaggle under Apache 2.0; Vertex AI Model Garden listing 'coming soon'

Gemini 4 Argon

A new Gemini 4 model that Google now presents as its frontier model. Google aims it at long software-engineering tasks, legal and finance knowledge work, and cyber defence. Access is limited to vetted security teams for now, with a phased rollout promised but not dated.

• Output limit of 1M tokens, up from 64K, per Google
• Long-horizon coding, multimodal understanding, vulnerability detection and patching
• Google-reported DeepSWE v1.1: 77.9%
• Google-reported LVBench long-video understanding: 91.7%
• Google-reported AutomationBench: 51.3% (#1)

AccessRestricted: vetted cyber defenders through Google's Fairwind Program only. Google says paid API customers and Google AI Ultra subscribers come next, then wider developer, enterprise and consumer access; no dates given.

Price$2 input / $10 output per 1M tokens introductory; $4 / $20 after the introductory period; cached input 95% off

Gemini 3.8 Live

Google's real-time voice model for spoken conversation, released with a 3.8 Live Extended Thinking variant that reasons while it speaks. It powers Gemini Live in the app and Search Live, and developers get it through the Live API. A Live Avatar option adds a lip-synced video persona for enterprise customers.

• Real-time speech conversation in 97 languages with mid-conversation language detection
• Runs tool calls in the background without pausing dialogue
• Extended Thinking: #1 on Artificial Analysis S2S Quality Index (82.6), per Google
• Google-reported 97.7% on Big Bench Audio

AccessGemini app (Gemini Live), Search Live, Gemini Live API and Google AI Studio; Vertex AI Studio private preview. A Live Avatar option is offered through Gemini Enterprise (post of 24 Sep 2026).

PriceText $0.75 in / $4.50 out; audio $3.00 in / $12.00 out per 1M tokens (Gemini API pricing page)

Gemini 3.8 Flash

The newest model in Google's Flash line, which Google calls its most intelligent workhorse model. It improves on 3.7 Flash in software engineering, agent tasks and multi-step reasoning. It is the most capable generally available Gemini text model in the app and API. A cybersecurity-tuned 3.8 Flash Cyber variant is restricted.

• 1,048,576-token input, 65,536-token output
• Text, image, video, audio and PDF in; text out
• Thinking levels, computer use (preview), Search and Maps grounding
• Google-reported HLE-Verified: 54.9%

AccessGemini app, AI Mode in Google Search, Google Sheets, Gemini API / Google AI Studio, Android Studio, Google Antigravity, Gemini Enterprise. The Gemini 3.8 Flash Cyber variant is restricted to Fairwind Program defenders (application required).

Price$0.75 input / $3.75 output per 1M tokens until 31 Dec 2026; $1.50 / $7.50 from 1 Jan 2027

Gemini Omni Flash

The first model in Google's Gemini Omni family, built to generate and edit video from any mix of text, image, audio and video inputs. It supports conversational, multi-turn edits that keep a scene consistent. An August 1.1 update added longer scene extension, 4K upscaling and cheaper 360p drafts.

• Video generation and editing from text, image, audio and video inputs
• Multi-turn natural-language edits that keep scene continuity
• 1.1 update: scene extension to 40 seconds, 4K upscaling, 360p drafts
• SynthID watermark on all outputs

AccessAt launch: Gemini app, Google Flow and YouTube Shorts / YouTube Create (Google AI Plus, Pro and Ultra subscribers). API public preview 30 Jun 2026; Gemini Omni 1.1 Flash (gemini-omni-1.1-flash) available in Google AI Studio and the Gemini Enterprise Agent Platform API from 27 Aug 2026.

Gemini 3.5 Flash

Launched at Google I/O 2026 as the first Gemini 3.5 model, aimed at agent work and coding. Google said it beat Gemini 3.1 Pro on hard coding and agent benchmarks while running about four times faster than other frontier models. The Gemini 3.5 Pro promised for the following month had not appeared by October.

• 1,048,576-token input, 65,536-token output
• Text, image, video, audio and PDF in; text out
• Google-reported Terminal-Bench 2.1: 76.2%
• Computer use (preview) in the Gemini API

AccessGemini app, AI Mode in Google Search, Gemini API (Google AI Studio, Android Studio), Google Antigravity, Gemini Enterprise Agent Platform, Gemini Enterprise.

Price$1.50 input / $9.00 output per 1M tokens

Gemma 4

Google's fourth generation of Gemma open-weights models, released in four sizes under the Apache 2.0 licence. They take text, images and video, with audio input on the small edge models, and are aimed at reasoning and agent workflows on local hardware. A laptop-sized 12B model followed in June.

• Sizes: E2B, E4B, 26B MoE, 31B Dense; 12B added 3 Jun 2026
• Context: 128K on edge models, up to 256K on larger models
• Text, image and video input; audio input on E2B and E4B
• 31B ranked #3 open model on the Arena AI text leaderboard, per Google

AccessOpen weights under Apache 2.0; download from Hugging Face, Kaggle or Ollama.

Gemini 3.1 Pro

An upgrade to Gemini 3 Pro that Google described as a smarter baseline for complex problem-solving. Google says it more than doubled 3 Pro's verified ARC-AGI-2 score. It is still labelled preview in the Gemini API and is the newest Pro model developers can use, since the announced 3.5 Pro has not shipped.

• 1,048,576-token input, 65,536-token output
• Text, image, video, audio and PDF in; text out
• Google-reported ARC-AGI-2 (verified): 77.1%
• Custom-tools variant for bash and tool-heavy workflows

AccessPreview: Gemini API, Google AI Studio, Gemini CLI, Google Antigravity, Android Studio, Vertex AI, Gemini Enterprise; Gemini app and NotebookLM for Google AI Pro and Ultra users.

Price$2.00 input / $12.00 output per 1M tokens (prompts up to 200K tokens)

Gemini 3 Flash

The fast, lower-cost model of the Gemini 3 generation, which replaced 2.5 Flash as the Gemini app's default. Google described it as Pro-grade reasoning at Flash-level speed and cost, at under a quarter of 3 Pro's price. It is still served as a preview model in the Gemini API.

• Google-reported GPQA Diamond 90.4%; SWE-bench Verified 78%
• About 3x faster than 2.5 Pro, per Google
• About 30% fewer tokens than 2.5 Pro on average, per Google
• Text, image, video and audio input

AccessDefault model in the Gemini app and in AI Mode in Search; Gemini API (preview), Google AI Studio, Gemini CLI, Google Antigravity, Android Studio, Vertex AI, Gemini Enterprise.

Price$0.50 input / $3.00 output per 1M tokens (audio input $1.00)

Nano Banana Pro (Gemini 3 Pro Image)

Google's image generation and editing model built on Gemini 3 Pro's reasoning. It is aimed at legible multilingual text in images, fine creative controls and consistency across many reference images, with output up to 4K. A cheaper Flash-based Nano Banana 2 followed in February 2026.

• Output up to 4K resolution
• Blends up to 14 images; keeps up to 5 people consistent
• Multilingual text rendering inside images
• Lighting, focus and colour-grading edits; SynthID watermark

AccessGemini app ('Create images' with the Thinking model), AI Mode in Search (US, AI Pro and Ultra), Workspace Slides and Vids, Gemini API, Google AI Studio, Vertex AI, Flow (Ultra). API model gemini-3-pro-image went GA on 28 May 2026.

Gemini 3 Pro

The first model of the Gemini 3 generation, launched as Google's most intelligent model with stronger reasoning and multimodal understanding. Google reported it at the top of the LMArena leaderboard, and it launched alongside Google's Antigravity agentic coding platform. Gemini 3.1 Pro has since replaced it.

• 1M-token context window
• Understands text, images, video, audio and code
• Google-reported LMArena 1501 Elo; GPQA Diamond 91.9%
• Google-reported Humanity's Last Exam: 37.5% (no tools)

AccessGemini app (all users), AI Mode in Search (Google AI Pro and Ultra), Gemini API (preview), Google AI Studio, Gemini CLI, Google Antigravity, Vertex AI, Gemini Enterprise, and third parties including Cursor, GitHub and JetBrains.