There are plenty of LLM providers to choose from now. I suspect most people choose one less deliberately than the comparison charts suggest. They start with whichever product reaches them first, then stay because their conversations, files, and habits are already there. The services they use for everything else matter too.
ChatGPT benefited enormously, in my view, from getting into people's hands early. For many people I speak to, ChatGPT effectively means generative AI. That familiarity is hard to compete with. I would expect a large user base to give OpenAI plenty of feedback to iterate on, and the product has grown well beyond a text box: documents, voice, and connections to external tools are now part of the experience.
Friends and family, both technical and non-technical, generally tell me they find ChatGPT approachable. They like the interface and interactive features, even though most still use little beyond basic chat. That is anecdotal, but it has shaped how I think about these products. A feature can help make a system feel accessible without becoming something everyone uses daily.
My rough mental map is ChatGPT for almost anything, Claude for programming, and Gemini for Google's ecosystem. Those are associations and habits, not hard boundaries. This is a reflection on what I actually use, not a definitive benchmark or a survey of every model available.
The more I move between them, the less convinced I am that intelligence alone will be the enduring moat in consumer AI. A capable model matters. So does the system that lets you do something useful with it.
ChatGPT is where I think things through
I think ChatGPT has the best general interface of the systems I regularly use. The interaction design feels considered. Perhaps that comes from having so many users and opportunities to improve it. Perhaps OpenAI just has a very good product and design team. I cannot separate those explanations from the outside, but I notice the result.
Several people have told me they stayed with ChatGPT after getting better answers from another provider. They found ChatGPT easier to use. That is a small sample, not market research, but it gets at the central issue: a better answer in one comparison does not necessarily make a better everyday product. Having to fight the interface every time you want something is a real cost.
I mostly use it for ideation and iteration. I talk through questions, develop half-formed thoughts, and ask about science, life, art, and nature. I also use it for documents, transcription, and work that benefits from keeping a conversation going. Occasionally I use connectors and other tools, but much of the value still comes from asking a question and working through the answer.
In my experience, ChatGPT manages context better than the other systems I regularly use. I mean the practical experience of carrying a discussion forward, retaining the relevant constraints, and letting me return to an idea without rebuilding the whole conversation. That is different from comparing advertised context-window sizes. A model being able to accept a lot of text does not tell me how well the product will use it.
I do not think it wins every specialist task. Coding, search, and logistics often send me elsewhere. But when I want accessible, conversational intelligence, ChatGPT is the system I reach for most naturally. The training supplies the underlying capability; the surrounding product makes it comfortable to use repeatedly. Both deserve credit.
Claude is where I build
Claude is my preferred coding system. It feels more like a precise, occasionally robotic collaborator than something trying to imitate human conversation. I appreciate that when I am working. I have also found it more robust against irrelevant or adversarial instructions, including prompt-injection attempts that try to redirect it through material it reads. That is my experience, not a security evaluation or a claim that it is immune.
I think a substantial part of Anthropic's appeal to developers comes from the harness. By that I mean the software around the model: the tools, permissions, context, and execution loop that allow it to work on a project. Claude Code can inspect a repository, edit files, run commands, and check its work, from a terminal or alongside an IDE workflow. That changes what I can ask it to do.
Third-party tools such as OpenCode offer a similar style of interaction with a choice of model providers. The model and the harness are separate choices, even when they arrive together. I see the development workflow moving towards agents handling more of the lifecycle: understanding a task, planning changes, implementing them, reviewing the result, and running checks. Autocomplete and isolated code generation are only a small part of that.
I pay for Claude because it has been the most consistent coding system in my work. I use it extensively for code generation, planning and executing implementation tasks, reviewing code, validating changes, and larger jobs across repositories. Consistency matters more to me than an impressive answer that I cannot reproduce on the next task.
I maintain internal benchmarks built around the coding tasks I care about. Claude currently meets my expectations for code generation best across those tests. They are useful for deciding what I want to use; they are not a comprehensive scientific ranking. I may write a separate post about the tests, what they measure, and where they fall short.
Claude also has habits that annoy me. It can be excessively verbose, and it overengineers. A small script can turn into an exercise in production-grade abstractions. Sometimes I need a temporary tool that does one thing and then gets thrown away. I do not want to maintain an architecture for it. I have had to constrain Claude with custom skills, instructions, and workflows to get implementations that are concise and appropriate to the job.
Perhaps that behaviour reflects its training or the demands of enterprise customers. That is speculation. Whatever the cause, I still prefer it for hobby projects, benchmarking systems, internal tools, and disposable software, provided I make the scope clear.
I have tried OpenAI's Codex tooling, including recent models, for the same kind of development work. Across my internal tests and normal use, Claude's coding harness has generally been slightly more consistent. That is my current experience of the complete systems. It does not establish that every Claude model is better than every model available in Codex.
Gemini makes me tolerate the interface
I use Google's AI products frequently for practical questions about the outside world: shopping research, web search, travel planning, opening and closing times, and places I need to find. Google has a natural advantage here. Search and Maps are already part of how I answer those questions, while Gmail, Docs, Slides, and the rest of Workspace give it useful context for productivity tasks.
These connections are real, although they are not all interchangeable. Gemini Apps can find places and directions through Maps and retrieve information from Gmail, Docs, and Drive through its Workspace connection. Gemini also works inside Workspace products, including Docs and Slides. Access and supported actions depend on the account, settings, and product surface. Owning the ecosystem does not mean every task is available through one chat.
I generally find Google's externally grounded answers reasonably trustworthy. When accuracy matters, I still check the source. A generated answer about an opening time can be wrong even when it is presented next to a convincing map.
My problem is the interface. In my opinion, Gemini's interface and interaction design are among the weakest of the major providers. I dislike the push towards in-chat animations, generated interfaces, and widgets. Too often they feel like a technical demonstration looking for a use case. I want strong intelligence, usability, predictable interaction, and reliable tool use. More things moving inside the chat do not automatically improve any of those.
Google's connector ecosystem could be one of its biggest advantages. Yet I have found that even its higher-end models can struggle when a task requires several tools or connectors to work together. Finding information in one place and using it correctly in another is where the surrounding system has to earn its keep. Access alone is not enough.
Some individual tools are excellent, though. Nano Banana is one of the strongest image-generation offerings I have used within a mainstream premium AI subscription. Google's current image-generation page distinguishes Nano Banana 2 from Nano Banana Pro, with the latter available through a paid-plan option to regenerate an image. My praise is for the Nano Banana tools I have used, not a claim that I have benchmarked every variant. The image results are a good reason to keep using a product whose general interface I dislike.
The economics interest me more than the chat apps
Chinese models deserve to be taken seriously. The argument about how far they trail American frontier models, and how quickly any gap is closing, can obscure a more useful question: what becomes possible when capable models are cheaper to run or available to deploy yourself?
Calling them all distillations would be wrong. DeepSeek-R1 includes smaller distilled releases trained using outputs from the larger model. DeepSeek-V3 and Kimi K2 use mixture-of-experts architectures, which activate only a subset of their parameters for each token. Distillation and sparse computation are different approaches to making capability more economical. Releasing weights is a separate decision about how others can access and deploy a model. None of those facts, alone, proves a lower total cost for every workload.
Open weights also do not mean identical licences or free hosting. DeepSeek-R1's repository describes MIT licensing and additional considerations for its distilled derivatives; Kimi K2 uses a modified MIT licence. These are concrete release examples, not descriptions of every model either company offers.
Among ordinary users around me, I see much less use of these products than ChatGPT, Claude, or Gemini. I cannot infer global adoption from that. In the developer conversations I encounter, the attraction is more often cheaper API access, self-hosting, automation, or long-running workloads. Those uses could be strategically important without replacing anyone's favourite consumer chat app.
I have tried DeepSeek and Kimi. Kimi has been reasonably capable for coding, but has not matched Claude in my use. DeepSeek has sometimes filled the role of a cheaper ChatGPT alternative, although I have generally found it less capable for my needs. Both are good enough to take seriously. Neither has become a major part of my daily workflow.
Censorship and political restrictions are another concern for me. If a system cannot discuss certain subjects freely or reliably, that reduces its usefulness, whatever its coding ability. I would not assume that every Chinese model, or every deployment of released weights, behaves identically. The restrictions of a hosted service and the behaviour of a model someone runs themselves need separate scrutiny. I want room to explore difficult questions, and censorship belongs in my assessment of a product.
I expect these models to keep improving. If access to compute and advanced hardware becomes less uneven over the next several years, the capability gap may narrow further. For now, my impression is that American frontier providers are advancing faster in the areas I use and delivering more useful consumer products. That is a judgement about my workflow, not a measured national ranking. My prediction is that open-weight Chinese models may be most disruptive in inexpensive inference, specialised deployments, coding, automation, and long-running agents.
Grok is still on my list to try properly
Grok has recently attracted more of my attention. Its connection to social media, particularly X, and its positioning around fewer conversational restrictions interest me. I perceive a looser alignment style, meaning a different approach to which requests it follows or refuses. That impression is not an objective measure of its intelligence or reliability.
I have not used it seriously for professional work or run it through my internal benchmarks. My instinct is to let the product mature before relying on it, although it may be unfair to judge it without giving it a proper trial. I may eventually evaluate it systematically. For now, it is something I am watching, with too little experience to offer a firm verdict.
The product is the whole system
My current arrangement is straightforward: ChatGPT for reasoning, conversation, ideation, transcription, documents, and context-rich interaction; Claude for software development; Gemini for search, shopping, travel, Maps, and Google-connected tasks. Chinese and other open-weight models remain promising lower-cost options I use less often. Grok is still waiting for a proper evaluation.
This arrangement is temporary. Models, interfaces, products, and prices change too quickly for permanent rankings. I have no reason to stay loyal to a provider if another system starts doing the work better.
What I expect to last is the importance of the complete system. We experience intelligence through an interface, context management, a coding harness, connectors, integrations, and all the small decisions that make an interaction reliable or frustrating. The provider with the smartest model on a benchmark may not be the one people keep using. The more defensible advantage may be making intelligence easy to access, direct, trust, and incorporate into everyday work. That is increasingly what I find myself paying for.