どのLLMを使うべきか？

•llmresearch-toolsmachine-learninglang:ja

研究で使うLLMを、ベンチマーク性能と費用・モデルサイズから比較します。点を選ぶと、モデル・reasoning effort・測定値・出典を確認できます。

MODEL QUALITY–COST FRONTIER

Which model setting gives the most capability for its measured cost?

Choose a benchmark below. The plotted model settings and both axes update together so that scores and costs always come from the same evaluation source. Upper-left is better.

Artificial Analysis v5.0Logarithmic cost axisUpdated 2026-11-01

Intelligence Index v5.0, reviewed 2026-11-01. Only published score/cost pairs from the same evaluation are shown.

Models without a comparable v5.0 result

GPT-5.6 Luna, GPT-5.6 Terra, GPT-5.6 Sol, GPT-6 Sol, Grok 4.6, Gemini 3.6 Flash, Gemini 3.1 Pro Preview.

Artificial Analysis Intelligence Index v5.0 versus weighted USD per task from the same evaluation. Each point is a published model / reasoning-effort setting.

GLM-5.3GPT-6 LunaGPT-6.1 SolGPT-6.1 TerraGPT-6 AstraClaude Opus 5.5Claude Sonnet 5.5Claude Fable 5.1Qwen3.8 MaxGemini 3.5 Flash-LiteLocal-model score reference
AA Intelligence Index versus measured evaluation costA logarithmic quality-versus-cost scatter plot using Artificial Analysis v5.0. Only model settings with a score and cost from the selected evaluation are plotted. In the Artificial Analysis view, a right-side rail also shows selected local-model scores without assigning them a cost coordinate. OpenAI markers are regular hexagons, Anthropic markers are triangles, and xAI markers are X shapes.$0.01$0.03$0.10$0.30$1.00$3.00$10.00$20.000102030405060Weighted cost per AA Intelligence Index task (USD, log scale)AA Intelligence IndexLOCAL REFERENCEQwen3.8 27B · 20LMxNLMHXMxLMHXMxNLMHXMxLMHXMxLMHXMxLMHXMxLMHXMxRR
N NoneL LowM MediumH HighR ReasoningX X-highMx Max
GLM-5.3 · Low
AA Intelligence Index23
Cost per Intelligence Index task$1.57
Plot coverage2 evaluated settings

Open weights · AA v5.0 measured sweep: Both published comparable v5.0 settings (Low and Max); High has no published AA score/cost pair. Custom GLM-5.3 license.

View evaluation source

Intelligence Index v5.0 is plotted against weighted USD per evaluation task. Lines connect every published comparable effort setting; missing settings are omitted. The local-model reference rail encodes score only.

AA current results AA methodology
LOCAL / OPEN-WEIGHT FRONTIER

How does open-weight model size compare with intelligence?

Current benchmark scores plotted against model size. The Artificial Analysis view uses v5.0, with every published setting for the selected families. Settings without a v5.0 score are omitted. The LiveBench view shows the saved snapshot and its sync status. Upper-left is better.

Artificial Analysis v5.0Local / open-weight candidatesReviewed 2026-11-01

Intelligence Index v5.0, reviewed 2026-11-01. Showing 21 models and 27 published settings. Large models include server-class references.

Tracked models without a v5.0 score

Qwen3.6 27B, Qwen3.6 35B-A3B, Gemma 4 26B A4B, Qwen3.5 9B, gpt-oss-20b, Gemma 4 E4B, Ministral 3 14B, Llama 4 Scout, Gemma 4 31B, Gemma 4 12B, gpt-oss-120b, Nemotron 3 Nano, Nemotron 3 Super, Nemotron 3 Nano Omni.

Current Artificial Analysis v5.0 scores versus total parameter count. Only published v5.0 scores are included; unavailable settings are omitted.

Select a model or focus a point to inspect its setting. All published points remain visible.

GLM-5.3Qwen3.8 27BMistral Small 4Muse GlimmerNemotron 3.5 LightningGranite 4.2 3BNemotron 3 Nano 4BNemotron 3.5 Super VLDeepSeek V4.1 Flash 1017MiMo-V2.6-FlashMiMo-V2.6-ProMiniMax-M3.1-FlashLing 3.1 Flash 1019Beam 1.0Muse Spark OpenStep 5 PreviewMistral Large 4K2 Horizon 375B A23BInkling SmallInklingGLM-5.3-FlashGPT-6 Astra score reference
Artificial Analysis Intelligence Index versus total parametersA logarithmic model-size scatter plot comparing local or open-weight LLM candidates. Artificial Analysis v5.0 and LiveBench overall views are selectable. AA-estimated values use dashed marker outlines. The Artificial Analysis views also include a right-side GPT-6 Astra score reference rail. OpenAI markers are regular hexagons; other local-model providers retain their circular markers.3B10B30B100B300B1000B2000B0102030405060Total parameters (billions, log scale)Artificial Analysis Intelligence IndexGPT-6 ASTRA REFERENCEL · 43M · 51H/X · 55Mx · 57GLM-5.3 · Low: 23GLM-5.3 · Max: 32Qwen3.8 27B · Non-reasoning: 10Qwen3.8 27B · Low: 16Qwen3.8 27B · Medium: 17Mistral Small 4 · Reasoning: 4Muse Glimmer · High: 6Nemotron 3.5 Lightning · Reasoning: 6Granite 4.2 3B · Reasoning: 5Nemotron 3 Nano 4B · Reasoning: 2Nemotron 3.5 Super VL · Reasoning: 14DeepSeek V4.1 Flash 1017 · Non-reasoning: 19DeepSeek V4.1 Flash 1017 · Max: 32MiMo-V2.6-Flash · Reasoning: 27MiMo-V2.6-Pro · Reasoning: 30MiniMax-M3.1-Flash · Reasoning: 29Ling 3.1 Flash 1019 · Reasoning: 29Beam 1.0 · Reasoning: 18Muse Spark Open · Xhigh: 25Step 5 Preview · Reasoning: 29Mistral Large 4 · Medium: 23Mistral Large 4 · High: 26K2 Horizon 375B A23B · Reasoning: 18Inkling Small · Reasoning: 13Inkling · Xhigh: 12GLM-5.3-Flash · Reasoning: 28Qwen3.8 27B · Xhigh: 20DQwen3.8 27B · Xhigh
D DenseM Mixture of ExpertsL Astra LowM Astra MediumH Astra HighX Astra X-highMx Astra Max
Qwen3.8 27B · Alibaba
Artificial Analysis Intelligence Index20
Total parameters27B
Architecture27B dense

AA setting: Xhigh · Reasoning · 262k context · Apache 2.0 · Open weights · LiveBench measured; AA v5.0 measured sweep: Dense 27B multimodal open-weight model. All four published v5.0 settings are included; High is not published. The LiveBench row comes from the dated saved snapshot.

View AA evaluation View model source

Total parameters describe the model-weight footprint; active parameters describe compute per token. Hundreds-of-billions and trillion-parameter models are server-class references. Practical memory requirements also depend on quantization, KV cache and runtime.

All published comparable v5.0 settings are shown. Vertical guides join settings at the same model size. The GPT-6 Astra rail encodes score only; equal displayed scores share a label.

AA open-weight results AA methodology

性能・費用図は、同じ評価で測定されたスコアとタスク費用を比較します。モデルサイズ図は、公開ウェイトモデルの総パラメータ数と1トークンあたりのactive parametersを切り替えられます。数百B〜1T級のモデルは、大規模サーバー向けの参照値も含みます。実行に必要なメモリは量子化・コンテキスト長・実行環境でも変わります。

AAはIntelligence Index v5.0の公開測定値に統一しています。LiveBenchとTerminal-Bench-Scienceは、それぞれ図に表示した日付・評価条件の保存データです。異なるベンチマークのスコアや費用を直接比較しないでください。

CC BY-NC 4.0 2026 © Shunichi Horigome.RSS