PD2V Web Which AI to use for what
AI changes almost every month. That's why this page has three parts. At the top is which AI to pick for which task — that changes slowly. Below that are the current versions and modes, with a date. At the bottom is my take on each AI and what I use myself.
🧭 Which AI for what
By task, not by version. This part holds even when the version numbers change.
How to instruct AI so the result is ready to use is in the How to write prompts guide.
🔢 Versions and modes
As of 24 October 2026. For hard tasks, turn on a mode that thinks longer. For current information, turn on web search.
| AI | Latest model | How to turn on thinking | Web search |
|---|---|---|---|
| Claude | Opus 5.5, plus Sonnet 5.5 and Haiku 5.5 | Click the model name next to the Send button, then Effort: Low, Medium, High, Extra high, Max | Menu + → Web search. In the new version of the app it searches on its own when it fits |
| ChatGPT | GPT-6 | In the model picker: Instant, Medium, High, Extra High, Pro (depending on the plan) | It searches on its own when it fits. You can turn it on by hand for a message |
| Gemini | Gemini 4 (Gemini 4 Argon in the API) | Model picker Flash-Lite, Flash, Pro and thinking level Low, Medium, High. Without a subscription, Auto only | It searches Google on its own |
| Grok | Grok 4.7 | Modes Auto, Fast, Expert, Heavy (Heavy only on the most expensive plan) | It searches the web and posts on X |
| Cursor | Composer 2.5, plus Claude, GPT, Gemini and Grok models | The thinking level is chosen with the model | Depends on the chosen model |
Midjourney is on version V8.1 (since 30 April 2026), DeepSeek on V4 (since 24 April 2026). Based on the makers' documentation and help pages. Names in the apps change often and differ by plan. If something doesn't match, check the model picker in your app.
💬 My take on each AI
From my own practice. The versions are as of the date above; opinions change more slowly.
Claude Opus 5.5
I do almost everything with Claude now: code, texts and analysis. I use it on the Extra High thinking level. It follows the brief precisely and writes the most human-sounding text of the big three (in Czech too, which is where I test it hardest). Besides the chat on claude.ai, it also runs in Claude Code (an agent in the terminal and editor) and in Cursor.
✅ Where it excels
- Coding: It finds its way around a big project, keeps the context across files and handles even a complex refactor.
- Writing: No corporate waffle and no clichés.
- Artifacts: Live HTML pages, charts and small tools right in the chat.
- Web search: It gives links to sources with every answer.
❌ What to watch out for
- Limits: On Extra High and Max you use up your plan's limits faster.
- It sometimes refuses even an innocent task. Explaining why you need it helps.
- It doesn't generate images or video. For images use ChatGPT, Gemini or Midjourney; for video, see AI video.
💡 What it's best for
Software development in Claude Code or Cursor, longer and sales texts, analysing contracts and documents, HTML overviews via Artifacts.
ChatGPT (GPT-6)
The most things in one app: chat, voice, images, Deep Research and an agent. Video was done by Sora, which shut down in April 2026. You choose how much it should think right in the model picker.
✅ Where it excels
- Voice mode: Smooth conversation, you can interrupt it and it sees through your phone's camera. Excellent for practising foreign languages.
- Images: It can do readable text in an image and edit parts of an image.
- Deep Research: It goes through dozens of websites by itself and returns a report with links.
- Data: It goes through Excel or CSV, finds errors and draws charts.
❌ What to watch out for
- “ChatGPT style”: It loves phrases like “fascinating” or “in today's world”. It sounds corporate unless you describe the style you want.
- Agrees too much: “Great question!” And when it's wrong, it doesn't like admitting the mistake.
- Top performance is only on the expensive Pro plan.
💡 What it's best for
Voice on your phone, analysing spreadsheets and PDFs, images for marketing, Deep Research for research.
Gemini 4
Gemini lives right inside Google: Gmail, Drive, Docs and YouTube. Under it are Nano Banana for editing photos and Gemini Omni for video, which replaced Veo in the app. It used to be my favourite; today Claude is.
✅ Where it excels
- YouTube: It summarises even an hour-long video and pulls out timestamps and a transcript.
- Lots of material at once: It takes in hundreds of PDF pages or a big chunk of code.
- Nano Banana: It keeps the face and style when editing a photo and can swap the background or clothes.
- Gemini Omni: A 10-second video with sound, from a description or from up to 5 photos. Requires a Google AI subscription.
- Google account: “Find the invoice in Gmail and summarise the cancellation terms from the PDF.”
❌ What to watch out for
- Cautious filters: It sometimes refuses an innocent thing without explanation.
- Writing: It sounds a touch more robotic than Claude. More of an analyst than a poet.
- Without a subscription there's only automatic model choice since 9 October 2026, usually the fastest Flash-Lite.
💡 What it's best for
Summaries of videos and lectures, big PDFs, photo edits with Nano Banana, ad videos with Gemini Omni, searching Gmail and Drive.
Grok 4.7
I use Grok alongside Claude, also on the Extra High level. Its biggest advantage is that it searches posts on X, so it knows what people are writing right now. It answers directly, with a bit of humour and without a corporate tone.
✅ Where it excels
- Live data from X: Trends, news, what people write about your brand. This is where it's strongest.
- Search with links: On the web and in posts on X.
- Direct tone: It can do sarcasm and humour and doesn't lecture.
- Code: xAI recommends Grok 4.7 for coding too.
- Images and video: Grok Imagine, only in paid SuperGrok. The prompts are in AI images.
❌ What to watch out for
- Images of people: In January 2026, Malaysia and Indonesia temporarily blocked Grok over sexualised deepfakes (the blocks have since been lifted), and authorities in the UK and California are investigating it. Don't edit photos of real people without their consent in any AI.
- It sometimes goes off the rails: It has looser settings. Check its output for business use.
- Writing: It writes a touch less naturally than Claude.
💡 What it's best for
What's happening right now and what people are saying about it, social media trends, research for trading, relaxed posts for social media. Prompts that make use of it are in How to write prompts and in the Prompt cookbook.
Cursor and Composer 2.5
Cursor is a code editor with a built-in agent. Its own model, Composer 2.5, is fast for everyday edits. But mainly, with one subscription you switch between Claude, GPT, Gemini and Grok.
✅ Where it excels
- Speed: Composer finishes an everyday change in a moment.
- All the big models in one place: You don't have to pay for each AI separately, and you get to try them in practice.
- It knows your project: It sees open files and the terminal and can run tests.
❌ What to watch out for
- Composer only lives in Cursor. You can't use it outside.
- For the hardest tasks (architecture, a big refactor), switch to Claude Opus.
💡 What it's best for
Adding a feature, fixing a bug, editing a file and tests in a loop. How to set up Cursor is in the Cursor guide.
Microsoft Copilot
Microsoft's AI on copilot.com, in Windows, on the Mac and on your phone. For free it does chat, images and working with uploaded files, plus Smart mode with GPT-5. With a Microsoft 365 subscription (Personal, Family, Premium) it's right inside Word, Excel, PowerPoint and Outlook.
✅ Where it excels
- Office: It summarises a Word document, suggests a formula or a chart in Excel and writes a reply in Outlook.
- Windows: It's built into the system, you don't have to install anything.
- Researcher and Analyst: In Microsoft 365 Premium it goes through sources and writes an overview, or breaks down numbers.
❌ What to watch out for
- Excel: The file has to be saved on OneDrive with AutoSave turned on.
- Family subscription: Only the subscription owner gets the AI features; they can't be shared.
- Deep Research for regular users ended on 18 August 2026; in Premium, Researcher replaces it.
💡 What it's best for
People who work in Word, Excel and Outlook and want AI right in the document. The prompts are in the cookbook → Office with Copilot. For coding there's a different product, GitHub Copilot, covered in GitHub for beginners.
Perplexity
Perplexity isn't a big model of its own but a search engine built on models from OpenAI, Anthropic, Google and its own Sonar. Every answer has links to sources. It also has its own browser, Comet, with an assistant.
✅ Where it excels
- Links with every answer: You see straight away where the AI got it from. Do click the citations though — even here the AI sometimes misreads a source.
- Deep Research: For buying decisions, competitor research or material for an article.
❌ What to watch out for
- It's not a generator: For long texts or code, use Claude or ChatGPT.
- It depends on the web: What isn't well described on the web, it won't find.
Midjourney V8.1
Still number one for artistic and stylised visuals. Version V8.1 has been out since 30 April 2026; it's faster and does 2K images without extra upscaling. It also does short videos.
✅ Where it excels
- Aesthetics: The most artistically coherent images.
- Your own style: Moodboards and style codes, so your brand has its own signature.
❌ What to watch out for
- Exact lettering and logos: Try ChatGPT or Gemini as well.
- Editing an existing photo while keeping the face works better in Nano Banana.
NotebookLM
It's not a typical chatbot. It's built on Gemini and works with the sources you upload: PDFs, course notes, your notes, links.
✅ Where it excels
- It answers from your sources and cites them. It makes things up much less than a regular chat. For important things, click the citation anyway.
- A podcast from your notes (Audio Overview): It turns a boring text into a conversation between two voices to listen to in the car or while cooking.
❌ What to watch out for
- It's not for free-form chatting: It sticks to the sources you give it.
💡 What it's best for
Preparing for final exams from your course notes, summaries of long PDFs and a podcast to listen to from your own work. More in the AI for students guide.
DeepSeek V4
A Chinese model with open weights. In January 2025 the R1 model shook the market; the current one is V4 (since 24 April 2026) in Pro and Flash versions. Through the API it's much cheaper than the American competition.
✅ Where it excels
- API price: Bulk text processing for a fraction of the price.
- Open weights: A company can run it on its own server and not send data out.
- Fewer Western filters for technical and creative tasks.
❌ What to watch out for
- Chinese censorship: The official chat refuses politically sensitive topics.
- Data: The official app and API run in China, so your texts go there.
- Writing works, but not as human-sounding as Claude.
🎯 My stack
🏆 What I use every day
- Main AI for everything: Claude Opus 5.5 on the Extra High level.
- Second AI: Grok 4.7 on the Extra High level.
For other tasks, see Which AI for what at the top. — PumpyDumpy