執筆に合うモデルを選ぶ
小説用モデルで重要なのは文章のリズムと質感です。一般やコードの評価は物語の力をほとんど示しません。声、テンポ、自然な会話、形式の遵守が原稿になじむかを決めます。
自分の文章でモデルを比べる
- 執筆には声やリズムと感覚描写、構成には因果の整合、設定検索には正確な記憶と統合が必要です。作業ごとに得意なモデルを選べます。
- 同じ場面、人物像、指示を2〜3モデルに送り、名前を見ずに読みます。残せる文章と書き直したい文章を記録すると、作家には評価点より役立ちます。
- まず自然な表現、語調、視点、会話の句読点、段落のリズムを見ます。本文だけを求めたとき、箇条書き、見出し、前置きを混ぜないか確認してください。
- 人物ごとの声、納得できる感情の衝突、世界のルール、先の展開の漏れを確認します。一部だけの修正を依頼して周囲の文章を保てるかも見てください。
- 待ち時間、メニューの消費倍率、手直しの量を合わせて考えます。推論を増やしても文章がよくなるとは限りません。複雑な構成には深い推論、普段の執筆には標準設定から試しましょう。
モデルの特徴と試し方
Based on the official model catalog (updated October 3, 2026). Suggested trials and notes are recommendations tailored for fiction writing rather than synthetic benchmarks; availability, versions, and usage multipliers follow your current App menu.
表を横にスクロールして全体を比較できます。
| モデル | 強みと試す方向 | 確認したい点 |
|---|---|---|
| Kimi K3公式資料 | Massive context capacity and deep information retrieval. Excellent for story bible lookups, cross-chapter foreshadowing verification, and evidence-grounded structural analysis. | Large context does not guarantee pinpoint recall on every subtle detail; monitor potential narrative drift and latency on long inputs. |
| GLM 5.3公式資料 | Engineered for complex reasoning and multi-threaded planning. Ideal for macro outline construction, cause-and-effect validation, and resolving character motivation deadlocks. | Reasoning steps add response latency; save it for deep brainstorming and structural reviews rather than quick, single-sentence polishes. |
| GLM 5.3 Flash公式資料 | Fast turnarounds and rigorous format compliance. Perfect as an agile writing assistant for scene beats, chapter title ideation, and rapid drafting passes. | Reasoning depth for sprawling multi-POV webs is more modest than the flagship; best deployed for agile drafting and sentence-level revisions. |
| Gemini 3.8 Flash公式資料 | Lightning-fast responses coupled with huge context windows. Great for broad-brush manuscript skimming, multi-angle premise generation, and interactive exploration. | Tone can lean clean and neutral in nuanced prose; provide concrete excerpt samples to prevent stylistic flattening. |
| DeepSeek V4.1 Flash公式資料 | Authentic narrative grounding, crisp concrete action verbs, and outstanding cost-efficiency. Offers both thinking and direct generation modes for flexible everyday drafting. | In thinking mode, monitor that analytical notes do not spill over into fictional prose; distinguish logical breakdown from storytelling subtext. |
| Qwen3.8 Flash公式資料 | Rich vernacular nuance, lively dialogue, and colorful idioms. Especially strong at period flavor, genre-specific vernacular, and grounded character interactions. | Give clear guidance on narrative restraint to prevent decorative vocabulary from overshadowing scene movement. |
| MiMo-V2.6-Flash公式資料 | Sub-second initial token latency and minimal resource footprint. Tailored for inline sentence completion, idiom replacement, and rhythm fine-tuning. | Excels at bounded local edits; not suited for managing sprawling chapter plans across a multi-volume epic. |
| Hy3公式資料 | Nuanced context comprehension and balanced emotional transitions. Reliable at steady progression and maintaining stable character dynamics across scenes. | Narrative voice defaults to a measured register; inject distinct stylistic excerpts when aiming for edgy or experimental prose. |
| GPT-6 Luna公式資料 | High-throughput, budget-friendly drafting. Well-suited for chapter titles, premise summaries, and quick scene staging with tight constraints. | May smooth over subtle psychological subtext; reserve climactic emotional turns for higher-tier reasoning partners. |
| GPT-6 Sol公式資料 | OpenAI flagship for sprawling narratives, intricate worldbuilding causality, and multi-faction conflict choreography. | Higher per-token rates and compute latency; deploy when resolving intricate contradictions and pivotal narrative twists. |
| Claude Opus 5.5公式資料 | The benchmark for literary aesthetic sensibility, subtext, and unspoken emotional tension. Remarkable restraint with virtually zero robotic sermonizing. | Higher latency and cost; best reserved as the lead editor for climactic revelations, layered relationship moments, and final manuscript polish. |
| Claude Sonnet 5.5公式資料 | Exceptional harmony of prose texture, narrative rhythm, and responsiveness. The gold standard for primary scene drafting and nuanced character dialogue. | Specify sections that must stay untouched to prevent it from over-smoothing intentional stylistic grit and raw vernacular. |
同じ文章を異なる方向に直すなら、複数の書き直し候補とブラインド選択で一案ずつ比べてから選べます。