Google Deep Mindプロプライエタリ

Gemini 3.5 Flash

Name: Gemini 3.5 Flash
Author: Google Deep Mind

このモデルを比較

Google DeepMind開発の最新マルチモーダルモデル。テキスト・画像・動画を統合処理。

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

2026-06-20

日本語性能

✅高品質日本語

多言語対応モデルのうち、日本語処理に優れた性能を持つモデル。

API料金

入力料金（1Mトークンあたり）

$1.5

出力料金（1Mトークンあたり）

課金モード: standard

強み

弱み

活用例

深度分析

Arena Elo

1479

#8 overall in Text Arena

SWE-Bench Verified

80.8%

per BenchLM aggregator; Google reports 55.1% on SWE-Bench Pro

Input Price

$1.50/1M

3x higher than Gemini 3 Flash; ~40% lower than Gemini 3.1 Pro

Output Speed

~284 tok/s

~4x faster than frontier peers

Context Window

1M tokens

Weak recall at full 1M (26.6% on MRCR)

Key Agentic Benchmark

83.6%

Field-leading MCP Atlas score

強み

・Best-in-class agentic tool use and multi-step workflow performance
・Exceptional speed (~4x faster than frontier models) at a competitive price
・Strong native multimodal capabilities including video and chart reasoning

弱み

・3x price increase over previous Flash generation breaks budget assumptions
・Poor long-context recall (26.6% at 1M tokens) despite the 1M window claim
・Lacks computer-use capability and trails on pure reasoning benchmarks vs. Pro

競合比較

Model	Arena	SWE	GPQA	Price
GPT-5.5 (OpenAI)	~1769	58.6% (Pro)	~92%	$5/$30 per 1M
Claude Opus 4.7 (Anthropic)	~1753	64.3% (Pro)	~91%	~$15/$75 per 1M
Gemini 3.1 Pro (Google)	~1314	54.2% (Pro)	94.3%	$2/$12 per 1M

概要

Gemini 3.5 Flash represents a strategic repositioning of Google's 'Flash' tier from a budget model to a frontier-class agentic specialist. Launched at Google I/O 2026, it delivers state-of-the-art performance in multi-step tool use (MCP Atlas: 83.6%) and agentic coding (Terminal-Bench 2.1: 76.2%), often beating the previous Gemini 3.1 Pro while running approximately 4x faster. This makes it the default production backbone for coding agents, computer automation, and multimodal workflows within the Google ecosystem. However, this upgrade comes with a significant 3x price increase over its predecessor, moving Flash into a mid-tier pricing bracket ($1.50/$9 per million tokens). While still 40% cheaper than Gemini 3.1 Pro, it is no longer the budget option. The model's key innovation is its thought preservation feature for multi-turn conversations and adjustable reasoning levels, but it has notable weaknesses: the 1M context window has poor recall at full length, it lacks computer-use support, and it trails Pro models on pure reasoning benchmarks like Humanity's Last Exam (40.2% vs. 44.4%). In summary, Gemini 3.5 Flash is the premier choice for high-volume, speed-sensitive agentic applications on the Google stack, but developers must recalibrate expectations and budgets away from the old Flash tier's economics and evaluate whether Pro-tier models are still needed for the hardest reasoning tasks.

ベンチマーク＆性能

Gemini 3.5 Flash delivers frontier-adjacent performance optimized for agentic and multimodal tasks, though with clear trade-offs in pure reasoning and long-context reliability. | Benchmark Category | Benchmark | Score | Context & Comparison | |-------------------|-----------|-------|----------------------| | **Agentic Tool Use** | MCP Atlas | **83.6%** | Field leader; beats Gemini 3.1 Pro (78.2%) and GPT-5.5 (75.3%). | | | Terminal-Bench 2.1 | 76.2% | Strong; 2nd behind GPT-5.5 (78.2%). | | | OSWorld-Verified | 78.4% | Competitive computer-use score, but note: model lacks computer-use capability. | | **Coding** | SWE-Bench Pro (Public) | 55.1% | Trails Claude Opus 4.7 (64.3%) and GPT-5.5 (58.6%). | | | SWE-Bench Verified | ~80.8% (aggregator) | Strong per aggregated reports, not on official card. | | **Multimodal Understanding** | MMMU-Pro | 83.6% | Field leader in official Flash class; near GPT-5.5 (81.2%). | | | CharXiv Reasoning | 84.2% | Excellent chart reasoning; beats Gemini 3.1 Pro (80.2%). | | **Reasoning & Knowledge** | Humanity's Last Exam | 40.2% | Behind Gemini 3.1 Pro (44.4%) and Claude Opus 4.7 (46.9%). | | | ARC-AGI-2 | 72.1% | Strong, but trails Gemini 3.1 Pro (77.1%) and GPT-5.5 (84.6%). | | | GPQA Diamond | ~92.2% | Excellent; trails only Gemini 3.1 Pro (94.3%). | | **Long Context** | MRCR v2 (128K) | 77.3% | Good, but behind Gemini 3.1 Pro (84.9%) and GPT-5.5 (94.8%). | | | MRCR v2 (1M) | 26.6% | **Poor recall**; unreliable at full window length. | | **Community Preference** | Arena Elo (Text) | 1479 | ~8th overall; strong for a mid-tier model. | | | Arena Elo (Code) | 1508 | Excellent in the coding category.

詳細比較

**vs. GPT-5.5 (OpenAI):** Gemini 3.5 Flash trades raw reasoning power and computer-use support for dramatically lower cost and higher speed. On MCP Atlas, Gemini leads (83.6% vs 75.3%), making it better for orchestrated tool workflows. On ARC-AGI-2 reasoning, GPT-5.5 wins decisively (84.6% vs 72.1%). At roughly 1/3 the input cost ($1.50 vs $5) and ~4x the speed, Gemini is the choice for cost-sensitive, high-volume agentic pipelines, while GPT-5.5 is for frontier reasoning and computer automation. **vs. Claude Opus 4.7 (Anthropic):** Claude remains superior for deep, pure coding (SWE-Bench Pro: 64.3% vs 55.1%) and creative/human judgment tasks. Gemini 3.5 Flash counters with massive cost savings (~1/10th the price), superior multimodal/chart understanding, and integrated Google Search grounding. For production agent fleets where cost and speed are critical, Gemini wins. For complex, single-shot software engineering or nuanced writing, Claude is superior. **vs. Gemini 3.1 Pro (Internal):** This is the intended intra-family split. 3.5 Flash beats 3.1 Pro on agentic/coding benchmarks while being ~40% cheaper and faster. However, 3.1 Pro retains a clear lead on pure reasoning (HLE: 44.4% vs 40.2%) and long-context recall (MRCR 128K: 84.9% vs 77.3%). The strategic choice: use 3.5 Flash as the default agent engine and escalate to 3.1 Pro only for the hardest reasoning problems until Gemini 3.5 Pro arrives.

コミュニティ評価

Developer and researcher reaction has been mixed but focused on its agentic strengths. On platforms like Arena.ai, it ranks highly in coding and hard-prompt categories (Coding Elo: 1508). Industry adopters like Box and JetBrains have praised its real-world performance boosts (e.g., 46.7% more accurate financial reports for Box, 10-20% better coding for JetBrains). However, the significant price increase from prior Flash models has drawn criticism, with commentators noting "Flash is no longer the budget tier." The weak 1M-context recall and ~61% measured hallucination rate have been flagged as serious limitations for factual, long-document work. The consensus is that it's a "legitimately fast" and "best-in-class for agents on Google's stack," but the price-performance trade-off requires careful calculation for teams migrating from older Flash versions.

ユースケース

1. **Production Coding & Agentic Workflows:** Ideal for teams running high-volume, multi-step coding agents (e.g., automated PR reviews, codebase refactoring loops). Its MCP Atlas leading score (83.6%) and speed make it reliable for complex tool orchestration. Choose it over GPT-5.5 or Claude when cost-per-task and loop latency are critical. 2. **Multimodal Business Analysis:** Perfect for applications requiring integrated analysis of charts, documents, and video (e.g., extracting data from financial reports, analyzing visual dashboards). Its CharXiv (84.2%) and MMMU-Pro (83.6%) scores are top-tier. Superior to text-only models for roles like financial analyst or market researcher. 3. **High-Volume, Low-Latency Customer Support:** The combination of 284 tok/s speed, cost-effective pricing, and Google Search grounding makes it excellent for real-time chatbots and support agents that need to pull current information. Better for this than slower, more expensive frontier models. 4. **Migration Path for Legacy Flash Users:** For teams still on Gemini 2.x or 3 Flash, it's the logical upgrade if they need significantly more capable agents, but they must budget for the ~3x cost increase. Not recommended if the use case (e.g., simple classification) didn't require the previous Flash tier's capabilities.