# Claude, ChatGPT and Gemini: what I use each one for Note, 2026-10-29 I use three AI environments: ChatGPT from OpenAI, Claude from Anthropic and Gemini from Google. None of them wins at everything. For everyday search and conversation I reach for ChatGPT, for sites and text I use Claude, and I open Gemini less often, mostly for Canvas. If someone doesn't program and has to pick one paid plan at around 100 PLN, I would point to Claude first, then ChatGPT, then Gemini. That is my opinion from my own use. I haven't tested Grok or Meta's models, so I won't say anything about them. Last updated: 2026-10-29 ## At a glance These are my subjective impressions from my own use. The details are below. | | ChatGPT | Claude | Gemini | | --- | --- | --- | --- | | What I use it for | Everyday search, voice conversation, Deep Research, programming in Codex | Sites, presentations and text, work in Claude Code | Google documents, learning in Canvas, video | | Pros | In my use I don't hit the chat limit, and Codex has its own; live voice mode; Sol as the best quality for the price; the cheap Luna for simple tasks, practically without a limit; images; control from the phone | Less mechanical language; better-looking sites and presentations; control from the phone; Sonnet is enough for an ordinary user | Built into Google Docs, Sheets and Slides; Canvas for learning; the only one of the three that generates video; Flash is very fast | | Cons | I prefer Claude for the look of sites and presentations; Tera didn't convince me | One limit for chat and Claude Code; no image or video generation of its own | Weaker models and a limit that runs out fast; Antigravity without SSH or phone control; Flash's quality is poor | ## ChatGPT: search and conversation ChatGPT is what I open most often. I look things up in it, analyse things and ask ordinary everyday questions. I also use Deep Research, a mode in which the assistant goes through many sources on its own and puts together a report. The main reason is limits. In the chat, in the browser and in the app, I don't hit any limit, and even if there is one, it doesn't affect Codex, which I use for programming, because that has its own. So conversations with the chat don't eat into the capacity I need for code. With Claude it is the other way round, more on that below. The second reason is the live voice mode. I talk to the assistant and it answers right away, instead of only turning my speech into text. The mode has a daily limit, which in practice I don't reach. Plain dictation into the microphone has no limit. I also add reminders and notes by voice, because ChatGPT has access to my tasks, notes and calendar (I described it on the [home page](https://majlo.dev/en/)). On the paid plan at around 100 PLN (in ChatGPT that is Plus) I get Sol, currently my favourite model. I think it offers the best quality for the price among the models I know. For tasks where accuracy matters more than style, I pick it. I saw it best in the [research on my plot](https://majlo.dev/en/blog/codex-research-dzialki): Sol is thorough and pulls data very carefully when I ask for it. Sonnet seems more superficial to me there, and Opus less eager to act. That is only my impression, not a test. I've seen similar differences at work too, but I won't describe those cases. OpenAI also has cheaper models: GPT-Luna, the cheapest, and GPT-Tera, sitting between Luna and Sol in price and quality. Luna is much worse than Sol, but not as weak as you might expect. In version 6.1 it is plenty for simple tasks, which it can grind through for as long as you like. It is very cheap, so on the plan at 100 PLN its use is, in my practice, practically endless. That is another plus for ChatGPT, and it convinces me. I used Tera for a while, in the GPT-5.6 version, and I won't go back: it burns a lot of tokens, like Sol, and the quality is lower than Sol's. That is my experience and opinion. I haven't tried version 6.1. Above Sol there is also GPT-Astra. It costs more than Sol and Opus, is priced close to Fable, and in my view it is the best model overall right now. Sol is somewhat weaker than it but much cheaper, which is why I use Sol day to day. I also use Codex for my own small apps, such as the plot and house planner. I build it so that I can control it through an agent instead of clicking around by hand. ## Claude: sites, presentations and text I use Claude mainly in Claude Code, among other things to build this site. I prefer its language to OpenAI's models: it sounds less mechanical to me. I dictate the texts for this site by voice and then run them through Humanizer, an extra set of instructions that teaches the agent to write less like an AI. I also prefer how Anthropic's models do visually: the frontend, meaning what the user sees on a page, and presentations look better to me than the competition's. This is how I split it. The house planner is only for me and doesn't have to look nice, it has to work, so I build it in Codex, because I think it does better technically there. When something has to look good and represent me, like a site for other people, I use Claude. The biggest minus is the limit. The browser, the app and Claude Code draw from the same pool, so if I ask something in the chat, I have less left for code. That is why I don't use Claude for everyday search and research. It irritates me, though I value its other strengths. Claude has no image or video generation of its own. It makes graphics with code, with what a browser can do, mostly JavaScript. Claude's mobile app also has a voice mode, but I don't use it. For conversation I prefer ChatGPT, because it has its own limit and I don't worry about burning the one I need for code. For an ordinary user who doesn't burn tokens on programming, Sonnet is enough in my view, and it is cheaper than Opus. On Claude the plan at around 100 PLN is Pro, and Opus is its strongest model. Fable only comes with the more expensive ones. Opus and Sol are cheaper than Fable and Astra and are enough for practically anyone. ## Control from the phone The Claude and ChatGPT apps have something Gemini doesn't: I can control from my phone an agent that runs on my computer. I have a [home server](https://majlo.dev/en/blog/agent-na-serwerze-domowym) where the agents run in server mode. I connect to them from my phone and they work on my projects without me getting up to the computer. This post came about that way too. I dictated it from my phone to Claude running on my computer, and it assembled the article on the site as I went. I looked at the local version, gave corrections, and only then deployed to the VPS, where the site is available to everyone. It works the same in ChatGPT. When I want Codex to fix something in a project, it takes a screenshot of how it looks, I can look at it and reply right away. I only get up to the computer when something fancier needs doing or I have to check something on the spot. ## Gemini: Google documents, Canvas and video Gemini has the Google AI Pro plan at around 100 PLN. The only reason that subscription would pay off for me today is the bundled YouTube Premium Lite. It sounds like a joke, but I think Google's current models are weaker than the competition. Argon is better than the previous ones, and yesterday, on 28 October, Carbon was announced, only ordinary users don't have it yet. The difference shows in the limits. Usage on these plans is counted in a five-hour window: once the limit runs out, you wait for it to renew. There is also a weekly limit, which renews every week. With my prompts Sol lasts two to three hours of the five-hour usage window. Argon stopped grinding after 30 minutes and did maybe a fifth, at most a quarter, of what Sol would do in the same window. That is my estimate by eye, not a measurement. The same goes for Opus: for the same money I can use Argon several times less. For programming Gemini is out. Google's environment, Antigravity, is weak as far as I'm concerned. It has two apps. The first, in agent mode, is similar to Claude Desktop or the ChatGPT desktop app and can't connect to a server over SSH. SSH is in Antigravity IDE, but there you have to look at the code, so for someone who only wants the result and doesn't look at code, it is awkward. It also can't be controlled from the phone. Deep Research in Gemini is fairly decent, but the limits make it hard to use. Gemini Flash is very fast, but the quality of its answers is poor in my view. Gemini does have strong sides, though. The most useful one is Google documents. If someone works a lot in Docs, Sheets and Slides, Gemini is built into them, can work on the document and answer questions about its content. I ran into a limitation here. With a presentation of over 50 slides, the built-in Gemini insisted it had 50 and couldn't find a slide around number 60. It was very irritating. I use it like this mostly at my job, so I don't know what the usage limit looks like there. For learning I use Canvas. When I'm learning something, I most often ask Gemini for a quick interactive page I can click through. I described how it looks in the [guide on learning with Gemini](https://majlo.dev/en/guides/nauka-z-gemini-notebooklm-canvas). ChatGPT has its own Canvas too, but I haven't used it, so I won't compare. Claude can create mini apps by itself, even without mentioning Canvas. Gemini also generates images and video. It is the only one of the three that generates video. OpenAI has shut down Sora, its video model, and Claude generates neither video nor images. ChatGPT and Gemini both have their own image models. If someone needs video, Gemini has it in one package with images and chat. An alternative is Higgsfield, a paid platform with many video models, including Seedance from ByteDance. You can hand it a prompt through MCP from Claude or ChatGPT. Google documents can also be connected to ChatGPT or Claude through MCP, so you don't have to stay with Gemini just to work in Google Workspace. ## Which one to pick - **You don't program and care about texts, sites and presentations**: Claude. Sonnet is plenty, and it wins on looks and language. - **You search a lot, do research, or Claude's shared limit bothers you**: ChatGPT. You don't hit a chat limit in my use, and you also get images and voice mode. - **You work in Google Docs and Sheets or need video**: Gemini. - **You program**: Codex or Claude Code. I don't recommend Gemini with Antigravity. For someone who doesn't program, I would rank them Claude, ChatGPT, Gemini. Someone who searches a lot and doesn't want to count limits can swap the first two. I've used Chinese models only for programming, so I'll write about them separately if I collect enough observations. ## Glossary - **ChatGPT**: An AI assistant from OpenAI that you talk to in a browser or an app, also by voice. - **OpenAI**: The company that makes ChatGPT and Codex. - **Claude**: An AI assistant from Anthropic, available in a browser, an app and as a tool for programmers. - **Anthropic**: The company that makes the Claude assistant. - **Gemini**: An AI assistant from Google, connected to your Google account and its apps. - **Canvas**: A view in Gemini where a page or document appears next to the chat, which you can run and refine. - **Deep Research**: A mode in which the assistant spends a few minutes going through many sources and writes a report from them. - **Codex**: An AI assistant from OpenAI that works on files on your computer: it reads them, changes them and runs programs. - **Sol**: One of OpenAI's AI models, available in ChatGPT. - **Claude Code**: A program from Anthropic in which the Claude assistant works on files and code on your computer. - **Humanizer**: An extra set of instructions (a "skill") for an AI agent that teaches it to write less artificially. - **frontend**: The part of a site or app the user sees in the browser, e.g. the looks, buttons and text. - **JavaScript**: A programming language that lets a page in the browser react to clicks and change without reloading. - **Sonnet**: One of Anthropic's AI models in Claude, cheaper than Opus. - **Opus**: One of Anthropic's AI models in Claude, the strongest one on the Pro plan at around 100 PLN. - **Fable**: One of Anthropic's AI models in Claude, available only on the more expensive plans. - **VPS**: A computer rented online where the site runs, with no need to keep it at home. - **YouTube Premium Lite**: A cheaper YouTube subscription that removes most ads when watching videos. - **Argon**: An AI model from Google, used in Gemini. - **Carbon**: A newer AI model from Google, announced on 28 October but not yet available to ordinary users. - **Antigravity**: Google's environment for programming with an AI agent, the counterpart of Codex and Claude Code. - **SSH**: A way to connect remotely to another computer, e.g. a home server, to work on it. - **GPT-Luna**: OpenAI's cheapest AI model. - **GPT-Tera**: An OpenAI AI model between GPT-Luna and Sol. - **GPT-Astra**: OpenAI's most expensive AI model, priced above Sol. - **Higgsfield**: A platform for generating video and images with various AI models, also through MCP. - **tokens**: Pieces of text that a model splits a conversation into. Their number decides how much you can use within a limit. - **prompt**: An instruction or question typed into an AI assistant. - **Google Workspace**: Google's suite of work apps: Gmail, Drive, Docs, Sheets and Calendar. - **Gemini Flash**: A fast AI model from Google in Gemini that answers instantly, but is weaker than the larger models. - **Sora**: OpenAI's model and app for generating video from a description, already shut down. - **Seedance**: An AI model from ByteDance that generates video from a description. - **MCP**: A standard that lets an AI assistant connect to outside apps, e.g. Google documents.