Which AI model shall we chose?

Published

January 12, 2026

In what follows, here is my personal take as of date:2026-01-12.

Basics

Generative AI based on Large Language Models (genAI) is great for many tasks. In this course we only focus on aspects of Data Analysis:

designing analysis
writing code
data wrangling such as joining tables, sample design, variable transformations
exploratory data analysis
modelling, regressions, machine learning
causal inference
creating tables and graphs
writing reports and presentations

Basics

Most models have a free and paid tiers. Free ones are good as kind of Google search replacement. For serious work you’ll be better off with paid tiers.
As of now the leading models are Google’s Gemini 3.0, OpenAI’s GPT-5.2, and Anthropic’s Claude Opus 4.5.
Each model have faster (cheaper) and thinking (more expensive) variants. For data analysis work, the thinking variants are recommended.
All major models have deep research
All major models support tool use (e.g. web browsing, code execution) and agentic patterns (multi-step workflows with memory and tool use).
Open weights models are also quite good, Deepseek and Q. You can even run a 7bn parameter model on an expensive computer

Flagship models: 2026 January 12

landscape overview

Feature	GPT-5.2	o3 (reasoning)	Gemini 3 Deep Think	Claude Opus 4.5
Design focus	Flagship multimodal with native reasoning	Deep reasoning for complex tasks	1M token context with extended thinking	Coding & complex instruction following
Multimodality	Text + images + audio + video (I/O)	Text + tool calls	Text + images + video	Text + images + Computer Use 2.0
Browsing / tools	Integrated search + MCP support	Auto-search for up-to-date facts	Grounded in Google Search	Deep integration with MCP & OS
Default style	Balanced, adaptive	Concise, source-cited	Thorough, exploratory	Professional, highly structured
Context window	400k tokens	128k tokens	1M tokens	500k tokens
Strengths	All-rounder, agentic workflows	Step-wise analysis, audit trail	Long-context research	Coding agents, OS automation
Weak spots	Expensive for simple tasks	No native media I/O	Slower response times	Slower than Haiku/Sonnet

Models for data analysis

For data-analysis projects, the recommended approach is:

Reasoning models (o3, Deep Think) for complex analysis requiring accuracy
GPT-5.2 for agentic workflows and multimodal tasks
Gemini 3 for long-context document analysis (entire datasets, reports)
Claude Opus 4.5 for coding and complex technical writing

All major models now support tool use (MCP) and agentic patterns.

Research & Writing

Synthesis Over Summarization: AI tools increasingly synthesize multi-source inputs into structured insights rather than paraphrasing single documents.
Security & Privacy: Modern workspaces rely on isolated execution contexts; strong non-training guarantees apply primarily to paid and enterprise tiers.
Multimodal Capability: AI can interpret charts, screenshots, and handwritten notes and incorporate them into drafts.

For data analysis workspaces – comparison

Workspace	2026 Key Features
Anthropic Claude Artifacts	• Creates interactive applications (tutors, calculators) within the output window. • Real-time iteration on complex document structures.
OpenAI ChatGPT Canvas	• Advanced frontier models with contextual persistence for tone and style. • Inline editing with granular control over specific sections.
Google NotebookLM	• Interactive Audio Overviews with user interruption and questioning. • Grounded citations linked directly to uploaded source segments.
Perplexity Pages	• Multi-source synthesis using live web retrieval. • Inline citation and consistency checking against sources.

Data Analysis details

Sandboxed Execution: Code runs in secure, ephemeral environments with no local system access.
Statistical Rigor: Strong support for Python-based libraries (e.g. pandas, scikit-learn) for exploratory and predictive analysis.
Direct Integration: AI can manipulate data directly within spreadsheets or dedicated analysis windows.
Limits: Reproducibility, package versions, and state persistence remain constrained relative to local workflows.

Data Analysis details

Workspace	2026 Key Features for Analysis
ChatGPT Data Analysis	• Executes Python in managed compute environments for multi-file datasets. • Assisted data cleaning and predictive modeling workflows.
Claude Analysis	• High-fidelity SVG and lightweight interactive output. • Fast iteration on statistical tables with publication-ready formatting.
Google Gemini in Sheets	• Multimodal cleaning: converts screenshots or PDFs into structured tables. • Natural-language formula generation and transformations.
Microsoft Copilot in Excel	• Native Python-in-Excel for statistical scripts inside spreadsheets. • Automated pivots, summaries, and forecasting via prompts.

Coding Assistance

Agent-Assisted Workflows: AI can coordinate multi-step tasks such as refactoring or bug fixing across large codebases, with human review.
Environment Security: Code is tested in secure sandboxes before changes are proposed.
Interconnected Tools: Integrated with development and collaboration platforms (e.g. Jira, Slack).

Coding Assistance — comparison

Workspace	2026 Key Features for Developers
Anthropic Claude Code	• Native VS Code extension with agent-style workflows and inline diffs. • Supports complex multi-file edits and testing assistance.
GitHub Copilot	• Uses Extensions to interact with external dev tools (Azure, Slack, Jira). • Deep context from local and remote repositories.
Cursor	• AI-first editor with awareness of project-wide dependencies. • Strong support for iterative refactors across files.
Windsurf (Codeium)	• Cascade mode for orchestrating large-scale refactors. • Robust free tier for students and individual users.

Security note: SOC2 compliance is common; strict zero-retention guarantees typically apply to enterprise or explicitly configured accounts.

What changed from 2025 Q2

No need to think much re which models to use.
Leading models similar capability, but different. Not really sure how…

Feedback

Dear Reader. I have limited experience. Suggestions are welcome, please post an issue.

--- title: "Which AI model shall we chose?" date: "2026-01-12" --- In what follows, here is my personal take as of `date:2026-01-12`. ## Basics Generative AI based on Large Language Models (genAI) is great for many tasks. In this course we only focus on aspects of Data Analysis: * designing analysis * writing code * data wrangling such as joining tables, sample design, variable transformations * exploratory data analysis * modelling, regressions, machine learning * causal inference * creating tables and graphs * writing reports and presentations ## Basics * Most models have a free and paid tiers. Free ones are good as kind of Google search replacement. For serious work you'll be better off with paid tiers. * As of now the leading models are Google's Gemini 3.0, OpenAI's GPT-5.2, and Anthropic's Claude Opus 4.5. * Each model have faster (cheaper) and thinking (more expensive) variants. For data analysis work, the thinking variants are recommended. * All major models have deep research * All major models support tool use (e.g. web browsing, code execution) and agentic patterns (multi-step workflows with memory and tool use). * Open weights models are also quite good, Deepseek and Q. You can even run a 7bn parameter model on an expensive computer # Flagship models: 2026 January 12 {.white background-color="#3b528b"} ## landscape overview | Feature | **GPT-5.2** | **o3** (reasoning) | **Gemini 3 Deep Think** | **Claude Opus 4.5** | |:--------------|:--------------|:--------------|:--------------|:--------------| | **Design focus** | Flagship multimodal with native reasoning | Deep reasoning for complex tasks | 1M token context with extended thinking | Coding & complex instruction following | | **Multimodality** | Text + images + audio + video (I/O) | Text + tool calls | Text + images + video | Text + images + Computer Use 2.0 | | **Browsing / tools** | Integrated search + MCP support | Auto-search for up-to-date facts | Grounded in Google Search | Deep integration with MCP & OS | | **Default style** | Balanced, adaptive | Concise, source-cited | Thorough, exploratory | Professional, highly structured | | **Context window** | 400k tokens | 128k tokens | 1M tokens | 500k tokens | | **Strengths** | All-rounder, agentic workflows | Step-wise analysis, audit trail | Long-context research | Coding agents, OS automation | | **Weak spots** | Expensive for simple tasks | No native media I/O | Slower response times | Slower than Haiku/Sonnet | ## Models for data analysis For data-analysis projects, the recommended approach is: - **Reasoning models (o3, Deep Think)** for complex analysis requiring accuracy - **GPT-5.2** for agentic workflows and multimodal tasks\ - **Gemini 3** for long-context document analysis (entire datasets, reports) - **Claude Opus 4.5** for coding and complex technical writing All major models now support tool use (MCP) and agentic patterns. ## Research & Writing {background-color="#f8f9fa"} * **Synthesis Over Summarization**: AI tools increasingly synthesize multi-source inputs into structured insights rather than paraphrasing single documents. * **Security & Privacy**: Modern workspaces rely on **isolated execution contexts**; strong non-training guarantees apply primarily to paid and enterprise tiers. * **Multimodal Capability**: AI can interpret charts, screenshots, and handwritten notes and incorporate them into drafts. ## For data analysis workspaces -- comparison | Workspace | 2026 Key Features | |:---|:---| | **Anthropic Claude Artifacts** | • Creates **interactive applications** (tutors, calculators) within the output window. • Real-time iteration on complex document structures. | | **OpenAI ChatGPT Canvas** | • Advanced frontier models with contextual persistence for tone and style. • Inline editing with granular control over specific sections. | | **Google NotebookLM** | • **Interactive Audio Overviews** with user interruption and questioning. • Grounded citations linked directly to uploaded source segments. | | **Perplexity Pages** | • Multi-source synthesis using live web retrieval. • Inline citation and consistency checking against sources. | ## Data Analysis details {background-color="#f8f9fa"} * **Sandboxed Execution**: Code runs in secure, ephemeral environments with no local system access. * **Statistical Rigor**: Strong support for Python-based libraries (e.g. pandas, scikit-learn) for exploratory and predictive analysis. * **Direct Integration**: AI can manipulate data directly within spreadsheets or dedicated analysis windows. * **Limits**: Reproducibility, package versions, and state persistence remain constrained relative to local workflows. ## Data Analysis details {background-color="#f8f9fa"} | Workspace | 2026 Key Features for Analysis | |:---|:---| | **ChatGPT Data Analysis** | • Executes Python in managed compute environments for multi-file datasets. • Assisted data cleaning and predictive modeling workflows. | | **Claude Analysis** | • High-fidelity **SVG and lightweight interactive output**. • Fast iteration on statistical tables with publication-ready formatting. | | **Google Gemini in Sheets** | • **Multimodal cleaning**: converts screenshots or PDFs into structured tables. • Natural-language formula generation and transformations. | | **Microsoft Copilot in Excel** | • **Native Python-in-Excel** for statistical scripts inside spreadsheets. • Automated pivots, summaries, and forecasting via prompts. | ## Coding Assistance {background-color="#f8f9fa"} * **Agent-Assisted Workflows**: AI can coordinate multi-step tasks such as refactoring or bug fixing across large codebases, with human review. * **Environment Security**: Code is tested in **secure sandboxes** before changes are proposed. * **Interconnected Tools**: Integrated with development and collaboration platforms (e.g. Jira, Slack). ## Coding Assistance --- comparison {background-color="#f8f9fa"} | Workspace | 2026 Key Features for Developers | |:---|:---| | **Anthropic Claude Code** | • Native **VS Code extension** with agent-style workflows and inline diffs. • Supports complex multi-file edits and testing assistance. | | **GitHub Copilot** | • Uses **Extensions** to interact with external dev tools (Azure, Slack, Jira). • Deep context from local and remote repositories. | | **Cursor** | • AI-first editor with awareness of project-wide dependencies. • Strong support for iterative refactors across files. | | **Windsurf (Codeium)** | • **Cascade mode** for orchestrating large-scale refactors. • Robust free tier for students and individual users. | **Security note:** SOC2 compliance is common; strict zero-retention guarantees typically apply to enterprise or explicitly configured accounts. ## What changed from 2025 Q2 * No need to think much re which models to use. * Leading models similar capability, but different. Not really sure how... ## Feedback Dear Reader. I have limited experience. Suggestions are welcome, please post an issue.