Use which model? This post has the answer.

Bundle of popular LLM models. Where are you?

What model do you have for each job?

.

We bring the model card or system card of each model to NotebookLM. Please give us the summary. We will give you the coordinates under the comment.

And here are text models, mainly coding, but in real life they can also be used.

-

🏠, what is each AI model?

- OpenAI: Home of ChatGPT, it will have any GPT model, such as GPT-5.2, GPT-5 mini, GPT-5.3-Codex, GPT-5.3 Codex Low, GPT-5.3 Codex High, GPT-5.3 Codex Extra High, GPT-5.3 Codex Fast, GPT-5.3 Codex Low Fast, GPT-5.3 Codex High Fast, GPT-5.3 Codex Extra High Fast, which also has a real sub-codex model.

- Anthropic: Claude's home, whether it's Claude Haiku, Claude Opus, Claude Sonnet.

- Google: This house will also be Gemini Flash with Gemini Pro.

- xAI: This one is from Elon Mask's X house, it will be Grok Code Fast 1, Grok 4.20.

Basically, we used the first three, and of course, if you can't imagine, choose auto.

This is in the IDE.

- Github Copilot: Well, it's Microsoft. There's a Raptor mini that Microsoft took the GPT-5 mini for fine-tuned, and Goldeneye took the GPT-5.1-Codex for fine-tuned. Both are used in vs code with Github Copilot specifically, and one place.

- Cursor: He himself has his own unique model as well. It's a Composer name. Composer 2.

-

🤖 what kind of job is each model suitable for Dave?

We threw each model card or system card of the model used in the github copilot to NotebookLM to analyze it, because the document of Github Copilot did not update the new model, so we took four topics there to analyze it.

.

⭐ Fast help with simple or repetitive tasks

Focus on simple tasks, do them quickly, focus on quantities such as text classification, introductory language translation, short retrieval, or general question-and-answer chat. There are small models like Gemini 3.1 Flash-Lite / 3 Flash, GPT-5 nano / GPT-5 mini, Claude Haiku 4.5.

.

⭐ General-purpose coding and writing

Focus on general use. Introduce Gemini 3 Pro, GPT-5.2, Claude Sonnet 4.6, a flagship model suitable for everyday use. Have a very good understanding of context. Can write people's language and computer language fluently and with few errors.

.

⭐ Deep reasoning and debugging.

Focus on complex applications and solve bugs. Introduce frontier models with an intensive thought process. Think before you answer, such as Gemini 3.1 Pro (with Deep Think Mode), GPT-5.4 pro / GPT-5.1-Codex-Max, Claude Opus 4.6.

.

⭐ Multi-model (screenshot, diagram)

Any work on images, diagrams, charts, or multi-models from screenshots or image files is analyzed.

Introduces multi-model models like Gemini 3 Pro / Flash, Claude Opus / Sonnet, GPT-5 that he understands the relationship between text and images very well.

.

✨ Gemini's distinctive features: Gemini's system supports uploading large files up to 2 GB. Whether it's video, audio files, images, or thick PDF documents, you can throw them in together to provide cross-modal AI analysis directly on the chat page.

-

🌐 for Web Developer

You can go to the Arena AI web (or the former LMArena) at Leaderboard in the Code category.

The latest update was on April 14, with over 1 million votes! and 334 models.

1. claude-opus-4-6-thinking

2. Claude-opus-4-6

3. Glm-5.1: This is the open source of the z.ai

4. Claude-sonnet-4-6

5. Claude-opus-4-5-2025 1101-thinking-32k

How do they score or vote? They will have a feature battle. We let AI do something like a cart checkout. They have 2 options for us to choose which one they like. We will not know which model each option comes from until we choose which one we like. The code is a one-page html that has stuffed everything.

-

📱 for Android Developer

The Android team has released Android Bench which model is most suitable for our Android app writing.

The latest update, April 7, adds new OpenAI models like GPT-5.4 and GPT-5.3-Codex.

1.GPT-5.4: 72.4% score

2.Gemini 3.1 Pro Preview: 72.4% score - same score as GPT-5.4, but slightly wider CI range value.

3.GPT-5.3-Codex: 67.7% score

4.Claude Opus 4.6: 66.6% score

5.GPT-5.2-Codex: 62.5% score

6.Claude Opus 4.5: 61.9% score

The score value is calculated based on an average of 10 rounds per 100 Test Cases.

Confidence Interval (CI) is the confidence score, the expected performance range of the model, based on the p-value statistically significant level < 0.0.

Since the score may be somewhat volatile, this CI allows us to know where the model's actual score will be.

So how does he think?

The problem comes from over 38,000 actual Pull Requests on Github, but 100 specific Android technical items such as Jetpack Compose, Coroutine with Flows, Room, Hilt are selected by writing a solution called a patch file.

In addition to running through pretty code, you have to run Unit Test or Instrumentation Test to get a score.

It's all about this. The link is under the comment.

# Includes IT matters # Share IT Trick # lemon 8 Howtoo # programmer # Developer