Google Deep MindProprietary

Gemini 3.1 Flash-Lite

Name: Gemini 3.1 Flash-Lite
Author: Google Deep Mind

Compare this model

A high-performance multimodal model developed by Google DeepMind for integrated processing of text, images, and video.

Share:X Hatena LINE

Parameters

Undisclosed

Context Window

License

Proprietary

Release Date

2026-05-07

Japanese Language Capability

✅High-Quality JP

Multilingual model with strong Japanese language processing capabilities.

API Pricing

API pricing for this model is not yet available

Strengths

Weaknesses

Use Cases

Deep Analysis

Release

March 3, 2026 (preview)

Cheapest, fastest Gemini 3-tier model; catalog GA 2026-05-07

Context Window

1,000,000 tokens

64K-65.5K token output ceiling; text/image/audio/video/PDF input, text output

Pricing

$0.25 / $1.50 per 1M

About 1/8 the cost of Gemini 3.1 Pro; no published cache discount

Speed

363 tokens/sec

TTFT ~2.5x faster than Gemini 2.5 Flash

GPQA Diamond

86.9%

vs GPT-5 mini 82.3%, Gemini 2.5 Flash 82.8% (no tools)

Strengths

・Best price-per-token in the Gemini 3 family at $0.25/$1.50, roughly 1/8 of Pro, for a 1M-token context and native multimodal input.
・Throughput king: 363 tokens/sec and a time-to-first-token about 2.5x faster than Gemini 2.5 Flash, built for high-volume, latency-sensitive jobs.
・Reasoning effort and reasoning budget controls let teams dial thinking down for classification/translation and up for harder tasks.
・GPQA Diamond 86.9% and MMMU-Pro 76.8% hold their own against models that cost far more.

Weaknesses

・Deliberately not an agent model - Google scopes it to data-processing workloads and publishes no agentic benchmark scores.
・Reasoning depth is capped versus 3.1 Pro; on the hardest long-context retrieval (MRCR 1M pointwise) it scores only ~12%, well behind Pro.
・Output speed and capability still trail frontier reasoning models; treat it as a workhorse, not a flagship.
・No open weights; only Google and approved partners serve it.

Competitor Comparison

Model	Arena	GPQA	Price
Gemini 3.1 Flash-Lite (this)	1432	86.9%	$0.25/$1.50
GPT-5 mini	-	82.3%	$1.00/$5.00
Claude 4.5 Haiku	-	84.3%	$1.00/$5.00
Gemini 2.5 Flash	-	82.8%	$0.30/$2.50

Overview

Gemini 3.1 Flash-Lite is Google DeepMind's cheapest, fastest Gemini 3 model, distilled from Gemini 3 Pro and built for high-volume, latency-sensitive work - translation, classification, and document pipelines - at $0.25/$1.50 per million tokens with a 1M-token context.

Benchmarks & Performance

On Google's own card it posts GPQA Diamond 86.9%, MMMU-Pro 76.8%, SimpleQA 43.3%, LiveCodeBench 72.0%, and MMMLU 88.9%, while hitting 363 tokens/sec. Arena Elo is 1432. It is explicitly positioned for data processing, not agent orchestration, and Google publishes no agentic benchmark.

Detailed Comparison

Against GPT-5 mini and Claude 4.5 Haiku it matches or beats GPQA Diamond at a fraction of the price; against Gemini 3.1 Pro it trades raw reasoning depth for ~8x lower cost and much higher throughput. The right comparison is the 'fast/cheap tier,' where it leads on speed.

Community Feedback

Developers on Reddit and Hacker News point to it as the default for bulk classification and RAG pre-processing; complaints center on capped reasoning and weak long-context retrieval at 1M tokens.

Use Cases

Best for translation at scale, content moderation, PDF/image batch extraction, and RAG answer generation where you need 1M-token context but can't pay Pro rates. Set reasoning budget low for simple tasks; reserve a frontier model for Terminal-Bench-class agentic work.

Latest News

Google DeepMind released Gemini 3.1 Flash-Lite on March 3, 2026 as a preview via AI Studio and Vertex AI, later listed as generally available around May 7, 2026. It anchors the 'brain + reflex' split with 3.1 Pro and is the cheapest entry in the Gemini 3 family at $0.25/$1.50 per million tokens.

Sources

Analysis generated: 2026-09-08