Googleの最速・最安モデルGemini 3.1 Flash-Lite、性能向上するも価格は3倍以上に
本文の状態
日本語全文を表示中
詳細モードで約1分の本文を読めます。
Google DeepmindがGemini 3シリーズ最速・最安モデル「Gemini 3.1 Flash-Lite」のプレビュー版を公開した。前モデルより大幅に高性能化されたが、出力コストは3倍以上に高騰している。
Source Article
元記事を日本語で読む
本文に関係しない購読案内、埋め込み通知、サイト内プロモーションは除いています。
Google Deepmindは、Gemini 3シリーズで最速かつ最安のモデルであるGemini 3.1 Flash-Liteのプレビュー版を公開しました。前身モデルよりも能力は大幅に向上していますが、出力コストは3倍以上に増加しています。
原文を表示
Google Deepmind has released a preview of Gemini 3.1 Flash-Lite, the fastest and most affordable model in the Gemini 3 series.
According to Artificial Analysis, the model scores 34 points on their Intelligence Index, a 12-point jump over its predecessor, Gemini 2.5 Flash-Lite. Despite the big leap in capability, it keeps up the same speed as the previous version, cranking out more than 360 tokens per second with an average response time of 5.1 seconds.
On multimodal tasks, it beats top-tier models like Claude Opus 4.6 and Kimi K2.5, hitting 78 percent on the MMMU-Pro benchmark. Artificial Analysis notes that tool usage barely improved. The context window stays at one million tokens.
On the Arena.ai leaderboard, which ranks models based on human preferences in blind comparisons, Gemini 3.1 Flash-Lite lands an Elo score of 1432. It outperforms other models in its weight class on reasoning and multimodal understanding, pulling 86.9 percent on GPQA Diamond (scientific knowledge) and 76.8 percent on MMMU Pro (multimodal understanding and reasoning). Those numbers top even larger previous-generation Gemini models, including 2.5 Flash.
Google says the model also delivers its first response token 2.5 times faster and pushes output 45 percent faster than Gemini 2.5 Flash (not 2.5 Flash-Lite; Flash is a larger model). Developers can dial how much the model "thinks," so it works for simple high-volume jobs like translations as well as heavier lifts like building user interfaces, according to Google.
The speed and quality gains come at a cost. Output pricing has more than tripled: Gemini 3.1 Flash-Lite charges $0.25 per million input tokens (up from $0.10 on the 2.5 version) and $1.50 per million output tokens (up from $0.40).
| Benchmark | Details of | Gemini 3.1 Flash-Lite (High) | Gemini 2.5 Flash (Dynamic) | Gemini 2.5 Flash-Lite (Dynamic) | GPT-5 mini (High) | Claude 4.5 Haiku (Extended Thinking) | Grok 4.1 Fast (Reasoning) |
|---|---|---|---|---|---|---|---|
| Input price ($/1M tokens, no caching) | Lower is better | $0.25 | $0.30 | $0.10 | $0.25 | $1.00 | $0.20 |
| Output price ($/1M tokens) | Lower is better | $1.50 | $2.50 | $0.40 | $2.00 | $5.00 | $0.50 |
| Output speed (Tokens/s) | 363 | 249 | 366 | 71 | 108 | 145 | |
| Humanity's Last Exam (Academic reasoning, full set, text + MM) | No tools | 16.0% | 11.0% | 6.9% | 16.7% | 9.7% | 17.6% |
| GPQA Diamond (Scientific knowledge) | No tools | 86.9% | 82.8% | 66.7% | 82.3% | 73.0% | 84.3% |
| MMMU-Pro (Multimodal understanding and reasoning) | No tools | 76.8% | 66.7% | 51.0% | 74.1% | 58.0% | 63.0% |
| CharXiv Reasoning (Information synthesis from complex charts) | 73.2% | 63.7% | 55.5% | 75.5% (+ python) | 61.7% | 31.6% | |
| Video-MMMU (Knowledge acquisition from videos) | 84.8% | 79.2% | 60.7% | 82.5% | - | 74.6% | |
| SimpleQA Verified (Parametric knowledge) | 43.3% | 28.1% | 11.5% | 9.5% | 5.5% | 19.5% | |
| FACTS Benchmark Suite (Factuality across grounding, parametric, search, and MM) | 40.6% | 50.4% | 17.9% | 33.7% | 18.6% | 42.1% | |
| MMMLU (Multilingual Q&A) | 88.9% | 86.6% | 84.5% | 84.9% | 83.0% | 86.8% | |
| LiveCodeBench (Code generation, UI: 1/1/2025-5/1/2025) | 72.0% | 62.6% | 34.3% | 80.4% | 53.2% | 76.5% | |
| MRCR v2 (8-needle) (Long context performance) | 128k (average) | 60.1% | 54.3% | 30.6% | 52.5% | 35.3% | 54.6% |
| 1M (pointwise) | 12.3% | 21.0% | 5.4% | Not Supported | Not Supported | 6.1% |
The model is available for testing in Google AI Studio and Vertex AI. Full benchmark results are up on Artificial Analysis and the Arena.ai Leaderboard.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
同じ出来事を4媒体で確認
同じ出来事を扱う別媒体の記事です。見出しと公開時刻を比較できます。
関連記事
News to Guide
ニュースの次に確認する
発表内容を、現在の料金や仕様と照らし合わせられる関連ガイドです。
今日のまとめ
AIデイリーブリーフで今日の重要ニュースをまとめ読み