Google Deep Mindプロプライエタリ

Gemini 3.5 Flash-Lite

Name: Gemini 3.5 Flash-Lite
Author: Google Deep Mind

このモデルを比較

Google Deep Mindのfoundationモデル。

シェア:X はてブ LINE

パラメータ

非公開

コンテキスト長

ライセンス

プロプライエタリ

リリース日

2026-07-21

API料金

入力料金（1Mトークンあたり）

$0.3

出力料金（1Mトークンあたり）

$2.5

課金モード: standard

強み

弱み

活用例

深度分析

Context Window

1,048,576 tokens

1M multimodal (text, image, audio, video, PDF)

Artificial Analysis Intelligence Index

#12 of 152 models; top of its class

Input Price

$0.30 / 1M tokens

Output: $2.50 / 1M; batch 50% off

Throughput

~490 tok/s

2nd-fastest measured on Artificial Analysis

SWE-Bench Pro

54.2%

vs Gemini 3.1 Flash-Lite 38%

Terminal-Bench 2.1

54.0%

vs Gemini 3.1 Flash-Lite 31%

OSWorld-Verified

74.0%

computer use; ~2x predecessor

強み

・Cheapest paid Gemini 3 tier at $0.30/$2.50 per 1M tokens, undercutting GPT-5.4 mini and Claude Haiku 4.5 on input price
・1M-token multimodal context with thinking, function calling, structured outputs, code execution and search grounding
・Near-top speed (~490 tok/s) makes it ideal for high-volume, latency-sensitive production traffic

弱み

・Loses to GPT-5.4 mini on sustained multi-step reasoning (Terminal-Bench 2.1, GDPval-AA v2)
・No computer use, image/audio generation, or Live API — a text-in/text-out engine for volume
・Time-to-first-token around 6s (includes thinking), so better suited to background jobs than snappy live chat

競合比較

Model	Arena	SWE	GPQA	Price
Gemini 3.5 Flash-Lite	N/A	54.2%	N/A	$0.30 / $2.50 per 1M
GPT-5.4 mini	N/A	N/A	N/A	$0.75 / $4.50 per 1M
Claude Haiku 4.5	N/A	N/A	N/A	$1.00 / $5.00 per 1M
Gemini 3.1 Flash-Lite	N/A	38%	N/A	prior budget tier

概要

Gemini 3.5 Flash-Lite is Google DeepMind's cheapest Gemini 3 model, released July 21, 2026 alongside Gemini 3.6 Flash for high-volume, latency-sensitive workloads such as translation, classification, and tagging. It carries a 1,048,576-token multimodal context (text, image, audio, video, PDF) at just $0.30 per million input and $2.50 per million output tokens, with a flat 50% discount in batch mode for non-real-time jobs. Google positions Flash-Lite as the layer beneath Gemini 3.6 Flash: 3.6 Flash for complex agent work, 3.5 Flash-Lite for the high-volume, latency-sensitive jobs that make up most production traffic — agentic search, document processing, translation, classification. On Artificial Analysis it lands an Intelligence Index of 36 (#12 of 152 models, well above the class average) while running at roughly 490 output tokens per second, the second-fastest model the service has measured. The generation jump over its March 2026 preview predecessor, Gemini 3.1 Flash-Lite, is the headline: SWE-Bench Pro nearly closes a 16-point gap (54.2% vs 38%), OSWorld-Verified climbs almost 20 points to 74.0%, and Terminal-Bench 2.1 doubles to 54.0%.

ベンチマーク＆性能

Gemini 3.5 Flash-Lite demonstrates strong, cost-focused performance across agentic and multimodal benchmarks, with most figures drawn from Google DeepMind's official Gemini 3.5 Flash-Lite model card. Treat cross-vendor columns as directional rather than neutral. | Benchmark | Gemini 3.5 Flash-Lite | GPT-5.4 mini | Claude Haiku 4.5 | | :--- | :--- | :--- | :--- | | **SWE-Bench Pro (Public)** | **54.2%** | higher | lower | | **Terminal-Bench 2.1** | 54.0% | higher | lower | | **MLE-Bench** | 39.2% | — | — | | **GDPval-AA v2 (Elo)** | 1140 | higher | higher | | **OSWorld-Verified** | **74.0%** | lower | lower | | **GDM-MRCR v2 (128K, avg)** | 72.2% | — | — | | **GDM-MRCR v2 (1M, pointwise)** | 21.3% | — | — | **Key takeaways:** * **Generational leap:** vs Gemini 3.1 Flash-Lite, SWE-Bench Pro nearly closes a 16-point gap, OSWorld-Verified climbs ~20 points, and knowledge work (GDPval-AA v2) jumps from 642 to 1140. * **Price leadership:** At $0.30 / $2.50 it undercuts GPT-5.4 mini ($0.75 / $4.50) and Claude Haiku 4.5 ($1.00 / $5.00) while sitting in the same rough quality band on most agentic benchmarks. * **Reasoning ceiling:** GPT-5.4 mini still wins on Terminal-Bench 2.1 and GDPval-AA v2, the benchmarks that reward sustained multi-step reasoning. * **Speed:** ~490 tok/s (2nd-fastest on Artificial Analysis); time-to-first-token ~6s including thinking, so better for background jobs than snappy live chat.

詳細比較

**vs. GPT-5.4 mini (OpenAI):** Flash-Lite undercuts it on price ($0.30/$2.50 vs $0.75/$4.50) and wins on computer use (OSWorld-Verified) and long-context recall (GDM-MRCR v2 128K), but loses on Terminal-Bench 2.1 and knowledge-work Elo where GPT-5.4 mini's stronger reasoning shows. **vs. Claude Haiku 4.5 (Anthropic):** Flash-Lite is ~3x cheaper on input and beats it on OSWorld-Verified and long-context retrieval, while Haiku holds a small edge on raw coding and reasoning scores. **vs. Gemini 3.1 Flash-Lite (predecessor):** The generation jump is clean — SWE-Bench Pro 54.2% vs 38%, OSWorld-Verified 74.0% vs ~54%, Terminal-Bench 2.1 54.0% vs 31%. Flash-Lite is the current cheapest stable option in the Gemini 3 family for teams migrating off Gemini 2.5 Flash-Lite (shutdown Oct 16, 2026). **vs. Gemini 3.6 Flash (sibling):** 3.6 Flash is for complex agentic work; 3.5 Flash-Lite is the high-volume, latency-sensitive workhorse. Google routes the two as a tier, not competitors.

コミュニティ評価

Developer reaction centers on its price-to-performance for volume work. Early adopters cited by Google include Ashler, Palo Alto Networks, and Ramp, all using it to scale agentic pipelines where per-call cost compounds. It is already rolling out inside Google Search's AI Mode. The main caveats raised in community discussions are the ~6s time-to-first-token (includes thinking), the lack of computer use / image-audio generation / Live API, and that the free tier trains on user content while paid tiers do not. The 50% batch discount ($0.15 / $1.25 per 1M) is widely noted as the right lever for overnight document processing and bulk translation.

ユースケース

1. **High-Volume Document Processing & Classification:** At $0.30/$2.50 with a 1M-token context, Flash-Lite is built for ingesting large PDFs, contracts, and support tickets at scale — e.g., first-pass triage on 50,000 tickets/month runs about $40 in raw token cost versus ~$150 on 3.6 Flash. 2. **Bulk Translation & Localization:** Multilingual translation across text, image, audio, and video inputs with batch mode at half price; ideal for overnight localization pipelines where latency is not the constraint. 3. **Agentic Search & Subagent Tooling:** As the layer beneath 3.6 Flash, it handles the millions of simple, repeated calls (retrieval, tagging, routing) that make up most production traffic, keeping per-call cost low while freeing the frontier model for hard cases. 4. **Long-Context Retrieval & Summarization:** GDM-MRCR v2 128K avg 72.2% makes it strong for needle-in-haystack recall over long documents, useful for RAG sidecars and meeting-note summarization.

最新ニュース

**Release (July 21, 2026):** Google shipped Gemini 3.5 Flash-Lite alongside Gemini 3.6 Flash and a gated Gemini 3.5 Flash Cyber security variant. Flash-Lite is the budget tier: $0.30/$2.50 per 1M tokens, undercutting GPT-5.4 mini and Claude Haiku 4.5 on input price. **Benchmarks & Positioning:** Google's model card shows SWE-Bench Pro 54.2%, Terminal-Bench 2.1 54.0%, OSWorld-Verified 74.0%, MLE-Bench 39.2%, GDPval-AA v2 Elo 1140. Intelligence Index 36 (#12 of 152) at ~490 tok/s. **Deployment:** Rolling out inside Google Search's AI Mode; named early adopters Ashler, Palo Alto Networks, and Ramp. Google's deprecation schedule still lists Gemini 3.1 Flash-Lite (not 3.5) as the migration target for Gemini 2.5 Flash-Lite, which shuts down October 16, 2026. **Pricing detail:** Standard $0.30/$2.50; batch and Flex $0.15/$1.25; Priority $0.54/$4.50; free tier trains on user content, paid tiers do not. Search grounding metered separately ($14 per 1,000 queries after 5,000 free/month).

Google positions Flash-Lite as the layer beneath Gemini 3.6 Flash: 3.6 Flash for complex agent work, 3.5 Flash-Lite for the high-volume, latency-sensitive jobs that make up most production traffic — agentic search, document processing, translation, classification. On Artificial Analysis it lands an Intelligence Index of 36 (#12 of 152 models, well above the class average) while running at roughly 490 output tokens per second, the second-fastest model the service has measured. The generation jump over its March 2026 preview predecessor, Gemini 3.1 Flash-Lite, is the headline: SWE-Bench Pro nearly closes a 16-point gap (54.2% vs 38%), OSWorld-Verified climbs almost 20 points to 74.0%, and Terminal-Bench 2.1 doubles to 54.0%.

出典

分析生成日: 2026-09-02