General-purpose assistant
GPT or Claude Sonnet
If you want to get moving without comparing for an hour, start here. These are the two simplest picks for thinking, writing and deciding.
See the frontier models →Living catalog · updated May 20, 2026
AI models:
which one to pick?
GPT, Claude, Gemini, Mistral, Llama, DeepSeek, Nano Banana, Sora, ElevenLabs... Everyone promises to be the best. Here, I sort by real use: writing, coding, searching, generating an image, making a video, keeping your data, or paying less.
Step 01
Start from what you need.
The most powerful one isn't always the best choice. The right model is the one that fits your task, your budget and your risk tolerance.
Code and agents
GPT-Codex, Claude, Qwen Coder
To edit a project, fix a bug or run tests, pick a model built to work with files and a terminal.
See the code models →Tight budget
Gemini Flash, DeepSeek, Qwen
When you handle a lot of short texts, the best model isn't always the priciest. The right pick is often the most consistent one.
See the low-cost models →Local / privacy
Qwen, Mistral, Phi, Gemma
Local mostly matters when your data can't leave for a provider. It isn't automatically simpler or cheaper.
See the open-weight models →Image
Nano Banana, Ideogram, FLUX
For visuals, I split the fast-draft models from the final-render ones.
Open the image page →Video
Gemini Omni, Sora, Veo
AI video isn't just an LLM with a camera. The criteria change: length, consistency, audio, editing.
See the video models →Simple guide
In your situation...
The right question isn't "which model is the strongest?". It's "what do I want it to do, with what data, and how many mistakes can I live with?".
Start with GPT or Claude Sonnet. They can write, summarize, compare, rephrase and help you decide without any fiddly settings.
Don't start with a local model: you'll be wrestling with the tech before you've even confirmed the use.
Claude Sonnet is the most natural pick for clean, nuanced, well-structured text. GPT is still excellent if you want more tools around it.
Before publishing, always ask for a shorter version and a more direct one.
Perplexity Sonar is made to answer with links and sources. It's closer to an answer engine than a creative assistant.
Keep one simple reflex: a cited source isn't necessarily a sufficient source.
Gemini is very strong when you hand it lots of PDFs, images, tables or transcripts all at once.
For a sensitive file with citations, Claude is still very good as long as the PDFs contain readable text.
GPT-Codex or Claude Code are the right picks when the AI has to read a project, edit files and run tests.
Don't ask it to "do everything." Ask for a plan, one small step, then a check.
Haiku, Gemini Flash, Qwen or DeepSeek are often more sensible than premium models for classifying, summarizing or extracting at scale.
The premium model is for the final check, not necessarily for every line you process.
Qwen, Mistral, Phi or Gemma can run locally or on your own server depending on the size you pick.
Local means more control, but also a machine, maintenance, logs and updates.
Don't just move down on price. Move down on difficulty: small models for simple tasks, big models for important decisions.
The bad trade is a cheaper model that makes you re-read everything twice.
ElevenLabs v3 or GPT-Realtime depending on whether you want a studio voice or a real-time conversation. For transcription, go with Whisper or Scribe.
If you want local without sending your voice to a provider: Moshi (Kyutai, CC-BY) is the only serious choice so far.
Claude Haiku 4.5 or Gemini 2.5 Flash-Lite are the right starting point. Small, fast, cheap, enough to answer 80% of customer questions.
Keep a premium model as backup for the tricky cases the small one can't handle.
| Word | Plain translation | What it changes for your choice |
|---|---|---|
| Context | Everything the model can keep in front of it while it answers. | Useful for big files. No point paying for very long if you ask short questions. |
| Open-weight | The model's weights are available, but the license can still be restricted. | Good for control and local use. Read the fine print before a commercial product. |
| RAG | The system first finds the right passages in your documents, then hands them to the model. | Better than one huge prompt if you have a document base that changes over time. |
| Reranking | A second, more precise sort after a broad search. | Very useful to stop the assistant from answering with the wrong document. |
Step 02
My shortcuts.
My top 6 right now. The ones I actually use every week — for my emails, my photos, my podcast, my in-house tools.
General assistant
Claude Sonnet 4.6
My default assistant for thinking, writing, organizing. The best price/quality balance when I want something clean on the first try.
llm
standard
closed
When I want the very best
GPT-5.4
When the task is complex or I can feel I'm going to spend time on it. Pricier, but often worth it.
llm
premium
closed
Photos & images
Nano Banana
What I use for my mom's Airbnb photos and the personal-branding photo generator. The best value for money on photo editing.
image
low cost
closed
Big documents
Gemini 2.5 Flash
For handling long PDFs, whole codebases or transcripts. A million tokens of context for under a dollar.
llm
low cost
closed
Voice & podcast
ElevenLabs v3
What I use for the voices on the Jerwis Productions podcast. Multilingual, multi-character, studio quality.
audio
standard
closed
Tight budget · open
Mistral Small 4
The savior of my high-volume scripts. Apache 2.0, so I can run it on my own machines if I need to. Unbeatable on price.
llm
low cost
open
Step 03
The catalog, by group.
52 models sorted by budget and by use. If you're unsure, look at My shortcuts above instead — that's my top 6.
Comparator
No model matches your search.
Try another keyword or look at "My shortcuts" above for the default picks.
Frontier · Premium
The most capable LLMs
When the result has to be top-tier: reasoning, complex code, agents, long contexts. Expect $15-30 per million tokens. You use these when the cost of a mistake outweighs the cost of the model.
Low cost · Volume
Great value for money
For handling lots of simple tasks: classifying, summarizing, extracting, tagging. Often 10× cheaper than the frontier models. Pair them with a premium model for the final check.
Open-weight · Local
The models you can host yourself
To keep control of your data or experiment without a subscription. Model size sets the hardware you need: 3B runs on a Mac, 70B needs a serious GPU.
Search · Sourced
When you search with sources
Optimized to answer with up-to-date links and citations. More of an answer engine than a creative assistant. Reflex to keep: a cited source isn't necessarily a sufficient source.
Image · Video · Audio
The creative models
Image, video and audio have their own page with a decision guide, a comparison table, a typical workflow and all the voices I use for the AI Wars podcast.
7 image models
Image
Nano Banana, GPT Image, Midjourney, Ideogram, FLUX, Stable Diffusion.
See the image page →
6 video models
Video
Sora, Veo 3.1, Gemini Omni, Runway Gen-4, Kling, Grok Imagine.
See the video page →
4 audio models
Audio
Eleven v3, Gemini Native Audio, Whisper, Voxtral. My podcast stack.
See the audio page →
RAG · Embeddings · Rerank
The building blocks for your document base
You don't use them on their own, you combine them to build a RAG: first you search your documents (embeddings), then you sort them (reranking), then you pass it to the LLM. More precise than pasting 100 pages into a prompt.
Platform-specific
Native built-in models
Models shipped straight by the OS or the platform. Worth knowing if you build on that base, otherwise you can skip them.
Not a magic ranking
I don't believe in one universal best model. A model can be excellent and still be a bad choice if the price, the latency or the data don't fit.
Open doesn't always mean free
I separate open source, open-weight, custom license and proprietary. It's less sexy, but it's what saves you from nasty surprises.
Monthly check
Each entry has a verification date. If a source changes, I fix it. If a model goes legacy, it moves to the archive.
Want the watch?
I test these models for my own projects. When a tool really deserves your attention, I put it in the newsletter.