Gemini vs ChatGPT vs Claude for data work: which to actually use

Every week someone asks me which AI they should use for their data work. They want one answer. I do not have one answer, because the honest truth is I use three, and I switch between them depending on the task like you switch between a screwdriver and a drill.

This is not a benchmark post. I do not care which model scores 0.3% higher on a leaderboard nobody runs their actual job on. This is which one I reach for, for which job, after a year of putting all three into a real BI workflow. No affiliate links, no hype.

The 30-second version

• Gemini: my default for SQL and anything I am automating. Free tier is generous, fast, and the 2.5 Flash model is more than enough for query generation and report narratives.
• ChatGPT: my pick for messy open-ended exploration, "here is a weird dataset, help me think." The code interpreter / data analysis mode is genuinely useful for poking at a CSV.
• Claude: my pick when the context is huge or the reasoning has to be careful. Big schemas, long docs, multi-step logic I cannot afford to get subtly wrong.

If you only take one thing: there is no winner. There is a right tool per task. Below is how I actually split them.

For writing SQL: Gemini

For day-to-day "write me this query," Gemini is my default and it is mostly about economics. It is fast, the free tier covers a working analyst easily, and for the kind of SQL most analysts write, the output is indistinguishable from the pricier options. When I automate a workflow that calls a model on every run, Gemini being free is the difference between a project that ships and one that needs a budget approval.

Caveat: any of the three will write good SQL if you feed it the schema and name the grain. The model matters less than the pattern. I wrote up that pattern separately, because using AI for SQL badly is worse than not using it at all.

For exploring a messy dataset: ChatGPT

When I have a CSV I do not understand yet, no clear question, just "what is even in here," ChatGPT's data analysis mode is the one I open. It will load the file, profile it, plot distributions, and let me poke at it conversationally. It is the closest thing to a junior analyst doing the boring first-pass profiling while I think about what actually matters.

It is not magic. It still makes the same grain and join mistakes as the others, and you still have to verify. But for the unstructured "help me orient" phase, it has the best ergonomics of the three.

For big schemas and careful reasoning: Claude

When I need to paste in a 40-table schema, a long PRD, three quarters of context, and have the model hold all of it without losing the thread, Claude is the one I trust most. The long-context behaviour is strong and it is less likely to confidently invent a column when the answer is genuinely "that is not in the schema you gave me."

I also reach for it on the analyses where being subtly wrong is expensive: anything going to leadership, anything with multi-step logic where one bad assumption poisons the conclusion. It tends to flag its own uncertainty more, which for an analyst is a feature, not a weakness.

What does not actually differ much

Honesty cuts both ways. For most everyday analyst tasks, the three are closer than the internet pretends. A 6-line report summary, a basic aggregation query, a "explain this error" question: all three handle these fine. If you are agonising over which to pick for routine work, you are optimising the wrong variable. Pick the one that is free and in front of you and move on.

The mistake that matters more than the model

Here is the thing nobody selling you a "best AI" listicle will say: which model you pick is the least important decision. How you prompt it is the most important. The same Gemini that hands you a wrong, fan-out query when you ask lazily will hand you a correct one when you feed it the schema, name the grain, and ask for the plan before the SQL.

I have watched analysts blame "Gemini is bad at SQL" when the real problem was they asked a model to write a query for a database it could not see. Switching to ChatGPT would not have fixed that. A better prompt would.

My actual stack, if you want to copy it

• Gemini (free AI Studio key) wired into my n8n automations and as my default SQL drafting tool.
• ChatGPT for the loose, exploratory CSV-poking sessions.
• Claude for the high-stakes, big-context, careful-reasoning work.

Total spend for most of this is zero, because Gemini's free tier carries the heavy automation load. You do not need a stack of paid subscriptions to add AI to an analyst job. You need one free key and the discipline to prompt it well.

Learn the prompting, not just the tool

In the 3-Hour Champion Sprint I teach the prompting patterns that make any of these three reliable against a real, messy Postgres warehouse, then wire Gemini into automated workflows you keep. The model is interchangeable. The skill is not. ₹1,499 Early-Bird, one Saturday, live.