Yes, they have their own personality
They aren’t human, but using them for a long time, fairly recognizable default styles emerge. However, a model’s behavior changes a lot depending on other factors, including: the specific model used, how we’ve set up the model’s various memories and instructions, the instructions within the single chat, various tools that alter the tone (e.g., Claude’s Styles), the skills loaded, and even the language of the conversation.
“Psychological” studies have been done on this that confirm there are differences, but net of the previous point, if we know well enough how to use them, ultimately we have the ability to largely modify their default characteristics to adapt them to the task we’re performing.
These assessments should therefore be taken as practical guidance, not as a verdict, and moreover, they are written having gathered data, but also and above all as my personal observations which, as such, are subjective, and will be presented with a certain degree of irony.
One last point: the usual caveat about the word “today,” meaning this is the situation as of July 2026, and it may well be different in a few months.
ChatGPT, the enthusiastic and capable all-rounder
ChatGPT tends to be sociable, fluent, and self-assured. Today it’s really very, very good at doing many things and doing them within a single interface, and even with code, the latest version of Codex has nothing to envy in Claude Code. It’s fun! You can easily talk with ChatGPT while driving, and the experience is anything but mechanical!
It’s very strong in operational reasoning, programming, step-by-step explanation, and quickly moving from idea to execution. The flip side is a tendency, in certain situations, to indulge the user too much. It can come across as overly encouraging, obliging, or a bit of a “yes-man”: phrases like “great idea,” “you’re doing great,” “you nailed it” can show up even when they’re not really needed, although lately this tendency toward enthusiastic sycophancy has been noticeably improved. In Milan we’d say it “doesn’t put on airs”; it has a very American approach, so enthusiastic, on top of things, versatile, very powerful, and at times a bit of a people-pleaser.
Claude, the ethical professional
It’s often more inclined than other models to say “I don’t know,” to flag uncertainties, and to correct the user when it believes a premise is wrong.
It’s probably the most professional, and it is so by design: Anthropic has focused its energies on making it an efficient tool for businesses, and it shows—in fact, it doesn’t do video, doesn’t do images, and even its voice system is rather poor, but that’s not something it cares about.
It’s more “European” in tone, even though it’s very American (its founder and CEO is Dario Amodei, of Italian origin, who left OpenAI over disagreements on safety-related issues—coincidence?).
Claude Code, its framework for code generation, was a revolution, and for quite a while, unmatched in performance; that’s no longer the case today, and other competitors are on par. Also worth mentioning is Cowork, which lets you act on your PC’s files and create scheduled tasks. In my opinion, it’s the one that seems “wisest,” meaning calmer and more attentive to deep ethical values, and this too is by design: Anthropic has founding principles that place particular emphasis on alignment and safety. It generally writes naturally, less mechanically, with a good argumentative rhythm. For a while it seemed overly cautious (read: boring); now it’s better.
Gemini, the super-connected one
Gemini tends to be structured, orderly, analytical. Its answers often feel more like a well-organized briefing than a natural conversation. It has some truly unique strengths! It’s integrated into the Google ecosystem, and if you use Workspace (look it up if you don’t know it), its potential is incredible—it can co-manage your email, your calendar, interact with Google Sheets, read and analyze the files in your Google Docs, etc. Very strong on multimodality, long context, search, extended documents, and price-to-performance ratio. Nano Banana, its built-in image generator, was a revolution.
Two more very important notes: YouTube is owned by Alphabet, which is Google, and its ability to accurately summarize video content is astonishing; moreover, Gemini is the LLM that powers NotebookLM, and if you don’t know what that is, you should immediately go look it up (of course, I’ll talk about it later in the course).
The problem is that many technical users, myself included, perceive a less stable quality compared to the initial peak of Gemini 2.5 Pro last year: sometimes it seems brilliant, other times flatter, colder, filtered, or inconsistent. On sensitive topics, it’s the politically-correct-and-a-bit-woke American of Silicon Valley; it can seem more locked down or less explicit, partly because the safety and filtering systems have a big impact on final behavior.
Grok, shall we grab a beer together?
A special mention for Elon Musk’s LLM: by far the friendliest.
Grok is probably underrated if judged solely by benchmark scores or its coding abilities, since it’s never quite at the level of the leading frontier models, but in pure chat it’s a blast! Tone, immediacy, irony, freedom of expression on every topic, and interaction with current events thanks to its strong connection with X (formerly Twitter): the guy has a big personality and, whatever people say, he’s not stupid at all!
Billing, on the other hand, is truly absurd: overlapping names, separate X and xAI plans, temporary offers, unstated consumer limits. And nothing, none of it makes sense—if you manage to figure it out, let me know.
DeepSeek, the pragmatist
DeepSeek is hard to describe if we’re talking about “personality”; it’s the least “characterful” among the major LLMs.
Still, it does have its own imprint: it’s dry, pragmatic, not very theatrical. It doesn’t have that American customer-service tone, always polite, reassuring, polished. It responds more like a technical engine than a personal assistant, and in practice it does a lot of work, at a price that’s not even comparable to the American frontier models.
And then there’s the paradox: it’s a Chinese model, and as such everyone points out that it’s “at the service of the Communist Party,” “who knows where our data ends up and how they’ll use it,” “if you ask it about Tiananmen Square it won’t answer you”—all true, but DeepSeek has had and continues to have a huge impact on the democratization of AI, because it’s a model that can be downloaded and run locally, with different sizes depending on the hardware available (the larger sizes need powerful servers, not home PCs); moreover, the parameters can be modified for fine-tuning, meaning making it specialized for particular tasks, and despite the tons of dollars and the cutting-edge chips used by American LLMs, the Chinese had to sharpen their thinking to make up for these shortcomings, and they succeeded by inventing extremely advanced technologies! While many major Western labs talk about safety, responsibility, and universal access, but keep models, data, costs, and infrastructure locked away, DeepSeek has put out into circulation models that are powerful, downloadable, modifiable, and usable commercially as well.
Mistral, the European alternative
I’ll admit I’ve used Mistral only a few times, but I remember them. The answers seemed to me totally neutral, even more so than DeepSeek’s; I tried it again just now and I confirm my personal impression.
Mistral is a family of French models, downloadable, accessible via the web through the Le Chat interface.
This interface, recently renamed Vibe, offers an excellent ecosystem of external connections: Google Workspace, Outlook, Slack, GitHub, Notion, Zapier, and custom MCP connectors.
Its strengths aren’t raw power, where it’s quite far from the frontier models, but the fact that it’s usable on a daily basis thanks to its connections (a big difference compared to DeepSeek, for example), that it’s downloadable and modifiable (the downloadable versions are free except for the top-tier ones, which are paid only for large businesses), and that it’s careful about data handling, being European and subject to the very strict GDPR.
Unfortunately, Le Chat, compared to the frontier models and given the performance it offers, is rather expensive.
There’s only one piece of advice: open up two or three you’ve never used and ask them the exact same question. The differences you’ll see are worth more than a thousand reviews.