Google Deep Mindプロプライエタリ

Gemini 3.6 Flash

Name: Gemini 3.6 Flash
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）

$1.5

出力料金（1Mトークンあたり）

$7.5

課金モード: standard

強み

弱み

活用例

深度分析

Context Window

1M tokens

1,048,576 input / 65,536 output; text, image, audio, video, PDF

Input/Output Price

$1.50/$7.50 per 1M

Standard API rate; ~1.7x cheaper than Gemini 3.1 Pro

SWE-Bench Pro

58.7%

+3.6 pts over Gemini 3.5 Flash (55.1%)

Terminal-Bench 2.1

78.0%

Agentic terminal workflows

OSWorld-Verified

83.0%

Computer use; +4.6 pts over 3.5 Flash

Token Efficiency

17% fewer output

vs Gemini 3.5 Flash on same tasks

強み

・Runs a full 1M-token multimodal context (text, image, audio, video, PDF) at Flash-tier latency and cost
・Strong agentic and tool-use performance — OSWorld-Verified 83.0%, Terminal-Bench 2.1 78.0%, MLE-Bench 63.9%
・About 17% fewer output tokens than 3.5 Flash with better coding, agentic, and knowledge-work scores; ~1.7x cheaper than 3.1 Pro

弱み

・Trails GPT-5.5 on Terminal-Bench (82.7% vs 78.0%) and Claude Opus 4.7 on OSWorld-Verified (82.8%) for heavy infrastructure work
・Reasoning ceilings below top reasoning models — HLE and ARC-AGI-style abstract tasks remain a gap versus Opus-class models
・Proprietary and Google-only; no open weights, and rate limits apply on high-volume agent loops without enterprise tiers

競合比較

Model	SWE	Price
Gemini 3.6 Flash	58.7% (SWE-Bench Pro)	$1.50/$7.50 per 1M
Gemini 3.5 Flash	55.1% (SWE-Bench Pro)	$1.50/$9.00 per 1M
GPT-5.5	58.6% (SWE-Bench Pro)	~$10/$30 per 1M
Claude Opus 4.7	64.3% (SWE-Bench Pro)	~$15/$75 per 1M

概要

Gemini 3.6 Flash is Google DeepMind's flagship efficient reasoning model, launched July 21, 2026. It pairs a 1,000,000-token multimodal context (text, image, audio, video, and PDF) with Flash-tier latency and a standard API price of $1.50 per million input and $7.50 per million output tokens. Google reports it uses about 17% fewer output tokens than Gemini 3.5 Flash on the same tasks while improving coding (SWE-Bench Pro 58.7%), agentic computer use (OSWorld-Verified 83.0%), and terminal work (Terminal-Bench 2.1 78.0%). It is the default Flash-tier workhorse for agentic coding, tool workflows, and long-document reasoning. Architecturally, 3.6 Flash continues Google's efficient-reasoning direction: strong tool calling, function calling, and computer use (screen understanding, UI actions) with fewer unwanted edit loops and hallucinations than 3.5 Flash. It leads the Flash family on MLE-Bench (63.9% vs 49.7%), CharXiv-R (89.4%), and MRCR v2 8-needle (54.0%), and holds a GDPval-AA agentic knowledge-work score around 1421. Against larger frontier models, it is roughly 1.7x cheaper than Gemini 3.1 Pro and far cheaper than Claude Opus 4.7 or GPT-5.5, while closing much of the coding and agentic gap. The main trade-offs are ceiling and access. On Terminal-Bench 2.1 it trails GPT-5.5 (82.7%) by about 4 points, and on OSWorld-Verified it trails Claude Opus 4.7 (82.8%) by a hair. Abstract-reasoning and HLE-style tasks remain below Opus-class models. As a proprietary, Google-only model with no open weights, it is best deployed through Vertex AI or the Gemini API for secure, scalable production agent fleets.

ベンチマーク＆性能

## Benchmark Comparison ### Coding & Software Engineering | Benchmark | Gemini 3.6 Flash | Gemini 3.5 Flash | GPT-5.5 | Claude Opus 4.7 | |---|---|---|---|---| | SWE-Bench Pro | 58.7% | 55.1% | 58.6% | 64.3% | | DeepSWE v1.1 | 49% | 37% | — | — | | MLE-Bench | 63.9% | 49.7% | — | — | | Terminal-Bench 2.1 | 78.0% | 76.2% | 82.7% | 70.0% | ### Agentic & Computer Use | Benchmark | Gemini 3.6 Flash | Gemini 3.5 Flash | GPT-5.5 | Claude Opus 4.7 | |---|---|---|---|---| | OSWorld-Verified | 83.0% | 78.4% | 78.7% | 82.8% | | MCP Atlas | 69.2% | — | 75.3% | 77.0% | | BrowseComp | — | — | 84.4% | 79.3% | ### Multimodal & Knowledge | Benchmark | Gemini 3.6 Flash | Gemini 3.5 Flash | |---|---|---| | MMMU-Pro | 84.3% | 80.5% | | CharXiv-R | 89.4% | — | | MRCR v2 (8-needle) | 54.0% | — | | GDPval-AA | ~1421 | ~1349 | ### Efficiency - Output tokens: ~17% fewer than 3.5 Flash on the same tasks - Context: 1,048,576 input / 65,536 output tokens - Latency: ~28 ms (vs ~41 ms for 3.1 Pro), throughput ~137 tokens/s - Cached inputs offer up to ~90% discounts

詳細比較

## Gemini 3.6 Flash vs Gemini 3.5 Flash 3.6 Flash is a strict upgrade over 3.5 Flash for agentic and coding work: +3.6 points on SWE-Bench Pro (58.7% vs 55.1%), +4.6 on OSWorld-Verified (83.0% vs 78.4%), and +1.8 on Terminal-Bench 2.1 (78.0% vs 76.2%). It also jumps MLE-Bench from 49.7% to 63.9% and produces ~17% fewer output tokens, lowering cost per task despite a slightly higher output rate ($7.50 vs $9.00 per 1M). For new production agents, 3.6 Flash is the recommended Flash tier. ## Gemini 3.6 Flash vs GPT-5.5 GPT-5.5 edges 3.6 Flash on Terminal-Bench 2.1 (82.7% vs 78.0%) and leads on strict interactive coding, but 3.6 Flash matches or beats it on SWE-Bench Pro (58.7% vs 58.6%) at roughly 1/7th the cost. For high-volume agent loops and multimodal document work, 3.6 Flash is the better value; escalate to GPT-5.5 only for shell-heavy infrastructure automation. ## Gemini 3.6 Flash vs Claude Opus 4.7 Opus 4.7 leads on OSWorld-Verified (82.8% vs 83.0% — essentially tied) and system-level reasoning, but costs ~10x more ($15/$75 vs $1.50/$7.50 per 1M). 3.6 Flash is the pragmatic default for day-to-day agentic coding, tool use, and long-context document analysis; reserve Opus 4.7 for the hardest architectural and reliability-critical tasks. ## Routing Recommendation - Coding agent / repo refactor / test repair → Gemini 3.6 Flash - High-stakes synthesis / complex architecture → Pro-tier (Opus 4.7 or GPT-5.5) - Classification / routing / tagging at massive scale → Gemini 3.5 Flash-Lite ($0.30/$2.50)

コミュニティ評価

Developer feedback on Gemini 3.6 Flash has been strongly positive for production agent workflows. Teams report fewer execution loops and hallucinations in agentic flows compared with 3.5 Flash, and praise its multimodal document handling (PDF, charts, tables, transcripts) and tool-use reliability. Enterprises are adopting it on Vertex AI for secure, scalable deployments, often as the default Flash-tier model behind agent frameworks. Community discussions highlight practical strengths over pure benchmark leaders: 3.6 Flash is favored for customer-support drafting, document synthesis with grounding, and computer-use automation where low latency and 1M context matter. Some note that for cyber/security variants, a 3.5 Flash Cyber pilot exists, and that 3.6 Flash is the safer drop-in replacement for legacy 3.5 Flash production flows after a regression pass. Independent comparisons (BenchLM, LLM-Stats) place 3.6 Flash's public score estimate around 75.3, overlapping with GPT-5.5 (72.28) within the 90% confidence interval, and note it as the cheaper, faster option for cache-heavy agent loops (roughly $0.135 vs $0.50 per loop at standard rates).

ユースケース

### 1. Production Coding Agents and Repository Refactoring Use Gemini 3.6 Flash as the default engine for coding agents that review pull requests, run security scans, and suggest tests. Its SWE-Bench Pro 58.7% and Terminal-Bench 2.1 78.0% scores, combined with ~17% fewer output tokens, keep per-task cost low at high volume. Example: a CI/CD agent that processes hundreds of PRs daily, generating fix patches and test suggestions at a fraction of frontier-model cost. ### 2. Multimodal Document and Knowledge Work With native PDF, image, audio, and video understanding plus a 1M-token context, 3.6 Flash ingests entire contract corpora, research papers with figures, or long meeting transcripts in a single prompt. CharXiv-R 89.4% and GDPval-AA ~1421 indicate strong chart/table and agentic knowledge-work capability. Example: due-diligence analysis over a 500K+ token legal corpus, or synthesizing a full ICLR paper including code and data. ### 3. Computer Use and Tool-Heavy Automation OSWorld-Verified 83.0% and strong function/tool calling make 3.6 Flash effective for UI automation, screen understanding, and multi-step tool workflows. It produces fewer unwanted edit loops than 3.5 Flash. Example: a back-office agent that navigates internal dashboards, fills forms, and reconciles data across systems. ### 4. Long-Context Customer Support and Search-Grounded Assistants The 1M context and low latency suit support answer drafting with grounding, transcript analysis, and retrieval-augmented research. Route low-confidence or simple classification tasks to Gemini 3.5 Flash-Lite for maximum cost efficiency, keeping 3.6 Flash for synthesis and evidence-bearing answers.