Claude
Anthropic
I lean on it for long documents and code that has to hold together.
The models I keep open.
I use all of these for writing, code, research and images. No single one is best at everything, and the differences start to matter once the work is real. These pages say where each one has an edge, and where it lets me down.
The tools
Pick the one you use
Gemini
Long context, the Google stack, and a quick image when I need one.
ChatGPT
OpenAI
The generalist I reach for first, and the easiest one to edit an image with.
Perplexity
Perplexity
Research with the sources attached.
DeepSeek
DeepSeek
Serious reasoning and code for pennies, as long as the data is not sensitive.
Grok
xAI
The one with live access to what people are saying right now.
Ollama
Open source
Models running on your own machine, with nothing leaving the room.
None of this is a boundary. Every model here has written code, drafted copy, answered questions and been asked for an image. What follows is where each one has an edge, not what it is allowed to do.
Also on the shelf
The rest of the shelf
Tools I use alongside the models above. Most of them are narrower than a general model, which is the point: they do one thing very well rather than everything adequately.
Llama
Meta
Open weights that run well locally, and the model most tools support first.
Qwen
Alibaba
Strong open models for code and for languages other than English.
Mistral
Mistral AI
Small, fast and European, which matters when a contract says the data stays in the EU.
Copilot
Microsoft
Useful inside the documents and spreadsheets a team already pays for.
Cursor
Anysphere
Writing code in the editor, with the whole project in view instead of one file.
Whisper
OpenAI
Turning a meeting recording into text I can search later. Runs locally too.
Midjourney
Midjourney
Still the best looking images, and the worst for anything that has to match a layout.
NotebookLM
Answers from documents I put in it, with the sentence it came from cited underneath.
LM Studio
LM Studio
A desktop window for local models when I do not want to touch a terminal.
ElevenLabs
ElevenLabs
Voice for anything that has to speak. Good enough that it needs a disclosure.
How I prompt all four
Four habits that do most of the work
Say who reads it
A note for a developer and a note for a client are different documents. One line about the reader changes the tone, the length and how much gets explained.
Paste a real example
One example of the output you want beats a paragraph describing it. That holds for writing, for code and for anything that has to match a format.
Ask for the shape
If the result goes into a spreadsheet or a program, name the fields and give one filled row. Models are good at filling a shape and poor at guessing one.
Make it show its work
Ask which sources it used, what it assumed and where it is unsure. You will not read that every time, but the answers get better when it knows you might.
Where I stop
What I do not hand to a model
The tools are good and getting better. These are still mine to decide.
- Whether the work is worth doing at all
- Spending money or signing anything on my behalf
- The final read before something reaches a client
- Judging a person, a CV or a performance review
Before you pick one
The questions I get asked most
Which model is actually the best?
There is no single answer, which is why I keep several open. Claude is stronger on long documents and careful code, Gemini on huge inputs and Google's tools, ChatGPT on everyday work and image edits, Perplexity on anything that needs a source, DeepSeek on heavy reasoning for very little money, Grok on what happened this week, and a local model when the data cannot leave the building. Every one of them can do the others' work. Those are the edges, not the limits.
Do you use open models or paid ones?
Both, for different reasons. Paid hosted models are stronger and I do not have to maintain them. Open models running locally are private, work offline and cost nothing per call. For anything with client data, where it runs matters more than which model wins a benchmark.
Is a cheaper model like DeepSeek safe for client work?
Only for work the client is happy to send there. The same goes for every hosted model, whatever it costs. When data cannot leave, I run an open model on a machine we control and accept that it is a little less capable. That is a conversation to have before anything is built, not after.
Do I need to pay for all of them?
No. Start with one paid plan on the model you use most, and use the free tiers for the rest. The gap between free and paid is mostly limits and speed, not a different model.
Can they make images good enough for client work?
For simple scenes and backgrounds, yes. For anything with text, a logo or a specific layout, I still finish the work in a design tool. A generated image is a starting point, not a deliverable.
How do you stop them inventing things?
By giving them the source and asking for the line it came from. For anything factual I use a tool that cites, then check the citation rather than the prose. A confident sentence is not evidence.
Would you build this into my team's workflow?
That is most of the work I do. The model is the small part. The rest is connecting it to your tools, deciding where a person approves something, and keeping the cost and the failures visible.
Let's build
Worth a thirty minute call?
Tell me what your team does by hand every week. I'll tell you honestly whether an agent is the right answer, what it would take, and roughly what it costs to run.