The Best AI Model Is the One You Can Actually Use
A personal note about comparing AI tools by whether they fit into real work, not just by how smart they seem.
Not a benchmark
This is not a benchmark. It is just what I noticed after using these tools for coding, projects, presentations, studying, and writing.
This month, somehow, I ended up with all of them: Gemini through a one-year student plan, GitHub Copilot Pro as a student, GPT Plus, and Claude Pro.
Gemini
Gemini is the easiest one for me to complain about.
For what I need, it has been pretty bad. The issue is not the brand or the interface. It is that when I try to use it for real work, I usually get something half-done. For coding, I trust GPT or Claude more. For project planning, Gemini feels less sharp. For presentations, the output is usually too shallow.
Maybe it is good at images or video. I do not want to be unfair there, because that is not my main use case. I mostly need these tools for coding, project work, presentations, and studying. In those areas, Gemini usually feels weaker than the others.
NotebookLM
NotebookLM is the exception.
It is probably the only part of Gemini that I really like. For exams that require memorization, it has genuinely helped me. The podcast feature is great because it turns study material into something I can listen to. That changes how I study, which is more than I can say for most AI features.
The video presentation feature was less useful. The results felt too short and too shallow. It might work as a quick overview, but not if I need something detailed.
Claude
Claude is a completely different story.
Claude Opus 4.7 is the best tool I have used for documents and presentations. It is organized, careful, and usually understands what I mean when I ask for changes. If I say “change these slides like this” or “make this section more structured”, it can actually do it.
That is the part I care about: Claude edits properly. It writes code, changes structure, and gives me something I can keep working on. NotebookLM sometimes gives presentation-like outputs too, but fixing the mistakes takes too much time. Claude is much easier to direct.
For presentations, Claude Opus 4.7 honestly feels like it is showing off. If I write a clear prompt, it usually gives me more than I asked for, in a good way.
But Claude has one huge problem: the limits.
I was excited to use Claude for coding too, especially because people kept praising Claude Code. But with Opus, two medium-hard tasks can be enough to hit the five-hour limit before the second one is even finished. If I use it heavily for a few days, I can also hit the weekly limit and then wait two days.
That is very frustrating.
The model is fast. The model is good. But I cannot use it freely enough. It feels like having a very capable teammate who disappears right when the work gets serious.
Coding in VS Code
For coding, I tried Claude, Copilot, and GPT inside VS Code. That is the comparison I care about, not random prompts in a chat window.
And honestly, I did not see a meaningful quality difference between Claude Opus 4.7 and GPT 5.5 for coding. Claude may feel smoother in some moments, but the latest GPT model is at the same level for my use. More importantly, GPT does not run out nearly as quickly. I can keep working.
That matters more than people admit.
ChatGPT is not great at presentations. At least for me, it does not come close to Claude there. But inside VS Code, with GPT 5.5, it is comfortable. It is reliable enough. It does not make me feel like I need to ration every serious request.
GitHub Copilot
GitHub Copilot is weird right now.
As a student, Copilot Pro sounds amazing. But recently, student accounts lost access to Opus models. Then, in the last week, newer GPT models after 5.2 also disappeared. That makes it hard to trust Copilot as the main tool. If the model I use today can just disappear from my plan, I cannot build my whole workflow around it.
Right now, I use Copilot more as part of my assistant setup. I might write about that separately after I spend more time with OpenClaw.
Context matters
The funny thing is that using all these tools made me care more about my own setup than the models themselves.
For agentic work, especially non-academic projects like this site or game projects, I started making small systems to save context. On this site, for example, I have an AGENTS.md file that tells coding agents the rules of the project: what the site is, which folders matter, what should not be changed, and what checks need to pass before a task is finished.
I also keep docs like ARCHITECTURE.md, docs/AGENT_MAP.md, and docs/WORKFLOW_REFERENCE.md. They are not fancy, but they stop the agent from guessing. Instead of rediscovering the same project structure every time, the agent can read the map and start closer to the actual problem.
That also saves usage.
And when every model has limits, context starts to feel expensive. A good project map can matter as much as a better prompt.
The end
So I do not think the question is “which AI model is the best?” anymore.
The best AI model is not always the smartest one. It is the one you can actually use when you need it.