A model for every idea
Explore AI models for research, writing, image editing and video generation. Compare measured performance and task costs to find a model for your work.
18 models
Research and writing
12- Claude Fable 5.1把研究摘錄、程式碼、日誌或流程圖交給 Claude Fable 5.1，整理證據、分析可能原因，把複雜問題拆成可以推進的計畫。
- Claude Haiku 5.5提供一段文字、一組簡短輸入或一個聚焦的問題給 Claude Haiku 5.5。將材料轉化為簡潔的摘要、實用的分類集或可與原始來源核對的細節清單。
- Claude Opus 4.6使用 Claude Opus 4.6 深入分析長篇資料，梳理複雜的程式碼變更，並細緻打磨草稿。提供相關背景與期望結果，再透過對話逐步完善內容。
- Claude Opus 5.5將想法與原始資料轉化為可用的成果。撰寫文章、分析決策、建立網頁或透過程式探索動態效果，再透過對話完善結果。
- Claude Sonnet 5.5從明確的目標和背後的素材開始。使用 Claude Sonnet 5.5 開發結構化文件、討論視覺參考或修訂草稿，同時保留您希望維持的事實和語氣。
- DeepSeek-V4.1-Flash將粗略的簡報、程式碼問題或一組筆記轉化為聚焦的草稿。定義任務，提供上下文，並請求下一步有用的操作。
- Gemini 3.8 Flash把參考資料整理成明確的下一步。使用 Gemini 3.8 Flash 梳理文件內容、比較圖片細節，或將概括的任務說明整理成結構清楚、可持續修改的草稿。
- Gemini 4 Argon用 Gemini 4 Argon 拆解數學問題、設計程式方案或閱讀技術長文。結合圖表與截圖，讓問題中的細節更容易說清楚。
- GPT-6.1 Sol把功能需求、程式碼變更或介面截圖交給 GPT-6.1 Sol，梳理出可以繼續推進的下一步。結合需求解釋複雜邏輯，再整理成實作計畫、測試案例或技術文件。
- GPT-6.1 Terra上傳圖片、一組筆記或未完成的草稿。運用 GPT-6.1 Terra 闡述關鍵細節、組織素材，並為目標讀者打造清晰的文字內容。
- GPT-6 Astra提出一個艱難的問題、背後的相關資料以及重要的限制條件。GPT-6 Astra 協助連結證據、權衡替代方案，並將結果轉化為可持續優化的清晰簡報。
- Grok 4.7用 Grok 4.7 審查程式碼、解釋截圖，或梳理難以取捨的技術方案。結合需求與相關資料，將討論整理成測試清單、遷移計畫或簡明的決策建議。
Create and edit images
4- GPT Image 2根據文字描述或參考圖製作產品圖、海報、人像與插畫。設定主體、風格與構圖，再調整重要細節。
- GPT Image 2.5 Sunburst改變場景，同時保留主體。當需要保留的細節與想修改的地方同樣重要時，GPT Image 2.5 Sunburst 可以協助你生成與調整圖片。
- GPT Image 2.6 Flare利用 GPT Image 2.6 Flare，將貼文構想或參考圖片轉化為圖片、可縮放 SVG 圖形或動畫 GIF。探索版面配置與插畫風格，並描述下一版本需調整之處。
- GPT Image 2.6 Sunburst使用 GPT Image 2.6 Sunburst 建立細節豐富的影像、可縮放 SVG 圖形與動態 GIF。描述主題、場景、光線與構圖，或上傳參考圖片以發展活動主視覺與文章配圖。
Bring ideas into motion
2Compare AI models
AI model research performance
Compare the factual accuracy of AI models when they research information and answer questions. Scores are out of 100; a higher score means more accurate answers on this set of tasks.
| Model | Low | Medium | High |
|---|---|---|---|
| 97.50 | 97.73 | No complete result | |
| 95.76 | 98.59 | 99.55 | |
GPT-6 Luna | 97.90 | 100.00 | 99.32 |
| 83.23 | 95.17 | 97.17 | |
| 91.46 | 96.18 | 97.40 | |
| 80.93 | 87.10 | 80.58 | |
| 89.92 | No complete result | 95.95 | |
| 0.00 | 43.07 | 65.36 |
GPT-6 Astra scored 99.55 for factual accuracy in the completed high-effort research evaluations shown here.
AI model writing performance
Compare how well AI writing models complete content creation tasks and deliver the requested results. Scores are out of 100 and measure task completion rather than article quality.
Original scoring criteria
| Model | Task fulfillment / 100 |
|---|---|
100.00 | |
100.00 | |
GPT-6 Luna | 99.00 |
99.00 | |
99.00 | |
99.00 | |
96.00 | |
95.00 | |
94.00 | |
93.00 |
Updated scoring criteria
AI model feature-use performance
Use tasks such as researching information, writing documents and creating slides to check whether AI models choose the right features and follow instructions. Scores measure feature use, not the quality of the finished articles, images or slides.
- Writing
- Research
- Images
- Slides
- Webpages
| Model | Feature-use score / 100Effort |
|---|---|
| |
| |
| |
| |
GPT-6 Luna |
|
| |
| |
|
Results with different scoring rules are shown in separate groups.
How to use
Choose your model
Choose an AI model for research, article writing or summarizing content. Switch between GPT, Claude, Gemini and other models in YouMind to explore different approaches.
Selected: GPT-6.1 Sol
Bring your context
Add PDFs, documents, images or notes, then describe your task. Ask AI to use your reference material to summarize key points, create an article outline or draft content.
Review and refine
Rewrite AI-generated text, refine the wording or select an image area for a local edit. Keep the parts you like and use specific instructions to refine the article or image.
Our Writing Sense analysis
Compare AI writing results and task costs
Compare Writing Sense scores with average agent task costs from the published report.
文感力與成本
In the published report's high-effort configurations, Claude Opus 5.5 scored 90.28 for Writing Sense at an average cost of $0.6035 per task. Gemini 3.8 Flash scored 78.85 at $0.1393 per task.
Higher effort did not always produce a higher score on these tasks. GPT-6.1 Sol scored 80.26 for Writing Sense at medium effort and 77.20 at high effort.
View data table
| Model | Effort | Writing Sense score / 100 | Understanding the author's intent / 100 | Following the author's intent in writing / 100 | USD / task |
|---|---|---|---|---|---|
| Claude Opus 5.5 | High | 90.28 | 100.00 | 87.04 | $0.6035 |
| GPT-6 Astra | High | 86.80 | 100.00 | 82.40 | $1.7338 |
| GPT-6.1 Sol | High | 77.20 | 92.59 | 72.07 | $0.4075 |
| Gemini 4 Argon | High | 81.72 | 100.00 | 75.63 | $0.1955 |
| Claude Sonnet 5.5 | High | 64.13 | 66.67 | 63.28 | $0.2787 |
| Claude Opus 4.6 | High | 63.11 | 77.78 | 58.22 | $0.4546 |
| Gemini 3.8 Flash | High | 78.85 | 77.78 | 79.21 | $0.1393 |
| DeepSeek V4.1 Flash | High | 59.96 | 62.96 | 58.96 | $0.0407 |
| GPT-6 Luna | High | 48.06 | 22.22 | 56.68 | $0.0646 |
| Claude Haiku 5.5 | High | 39.31 | 33.33 | 41.31 | $0.0227 |
| GPT-6.1 Terra | High | 73.15 | 77.78 | 71.61 | No comparable cost |
Which AI models understand your source material?
Compare error rates and the choices behind them.
Over-cuttingPolished but wrongOther
| Model | Reading error rate | Among wrong answers |
|---|---|---|
0.0% | No errors in this sample | |
4.9% | Over-cutting: 25.0%Polished but wrong: 75.0%Other: 0.0% | |
6.2% | Over-cutting: 80.0%Polished but wrong: 20.0%Other: 0.0% | |
8.6% | Over-cutting: 14.3%Polished but wrong: 71.4%Other: 14.3% | |
22.2% | Over-cutting: 50.0%Polished but wrong: 0.0%Other: 50.0% | |
23.5% | Over-cutting: 47.4%Polished but wrong: 5.3%Other: 47.4% | |
38.3% | Over-cutting: 67.7%Polished but wrong: 16.1%Other: 12.9% | |
45.7% | Over-cutting: 62.2%Polished but wrong: 13.5%Other: 24.3% | |
71.6% | Over-cutting: 19.0%Polished but wrong: 62.1%Other: 19.0% | |
75.3% | Over-cutting: 26.2%Polished but wrong: 39.3%Other: 32.8% |
View data table
| Model | Error rate | Over-cutting | Polished but wrong | Other |
|---|---|---|---|---|
| Gemini 4 Argon | 0.0% | Not applicable | Not applicable | Not applicable |
| GPT-6 Astra | 4.9% | 25.0% | 75.0% | 0.0% |
| Claude Opus 5.5 | 6.2% | 80.0% | 20.0% | 0.0% |
| GPT-6.1 Sol | 8.6% | 14.3% | 71.4% | 14.3% |
| Claude Opus 4.6 | 22.2% | 50.0% | 0.0% | 50.0% |
| Gemini 3.8 Flash | 23.5% | 47.4% | 5.3% | 47.4% |
| Claude Sonnet 5.5 | 38.3% | 67.7% | 16.1% | 12.9% |
| DeepSeek V4.1 Flash | 45.7% | 62.2% | 13.5% | 24.3% |
| GPT-6 Luna | 71.6% | 19.0% | 62.1% | 19.0% |
| Claude Haiku 5.5 | 75.3% | 26.2% | 39.3% | 32.8% |
How much do AI models write, and do they use the right material?
Can AI models write what you mean after understanding it?
Does reading well mean writing well? Compare model rankings in both tasks.
| Model | Reading | Writing |
|---|---|---|
Rank#1100.0% | Rank#373.2% | |
Rank#295.1% | Rank#275.5% | |
Rank#393.8% | Rank#184.9% | |
Rank#491.4% | Rank#572.0% | |
Rank#577.8% | Rank#757.5% | |
Rank#676.5% | Rank#472.7% | |
Rank#761.7% | Rank#664.1% | |
Rank#854.3% | Rank#855.8% | |
Rank#928.4% | Rank#954.4% | |
Rank#1024.7% | Rank#1048.6% |
When correct reading still leads to a wrong turn
Among the articles a model read correctly every time, how many still drifted from the author’s intent when rewritten?
View data table
| Model | Reading | Writing | Breaks an author choice |
|---|---|---|---|
| Claude Opus 4.6 | 77.8% | 57.5% | 66.7% |
| Claude Haiku 5.5 | 24.7% | 48.6% | No qualifying data |
| Claude Opus 5.5 | 93.8% | 84.9% | 44.4% |
| Claude Sonnet 5.5 | 61.7% | 64.1% | 85.2% |
| DeepSeek V4.1 Flash | 54.3% | 55.8% | 77.8% |
| Gemini 3.8 Flash | 76.5% | 72.7% | 46.3% |
| Gemini 4 Argon | 100.0% | 73.2% | 59.3% |
| GPT-6.1 Sol | 91.4% | 72.0% | 65.1% |
| GPT-6 Astra | 95.1% | 75.5% | 54.0% |
| GPT-6 Luna | 28.4% | 54.4% | No qualifying data |
Make the vase blue.