[Definitive Edition] How to use Gemini, Claude, and OpenAI as of mid-January 2025

Recently, for certain reasons, I subscribed to Google One—or whatever it's called—for Gemini Advanced. I'm going to use it to the fullest during the free trial period. I'm still debating whether to continue the subscription.

Furthermore, although a bit late, I subscribed to ChatGPT Pro. For certain reasons, I wanted to use it for personal purposes, so it's a personal subscription.

I was already subscribed to Claude Professional and ChatGPT Plus. I also have a subscription to Perplexity, but that's free for a year as a bonus for that piece-of-junk gadget, the Rabbit R1.

Recently, I've been making full use of conversational AI for writing and Google AI Hackathon projects, so I've accumulated a broad range of knowledge about the latest models and would like to compare them.

Since there are many services I've only just started using, I would appreciate it if you could provide any differing opinions or counterarguments in the comments.

Note: Thank you for the comments. I haven't used the Gemini series enough to fully realize its true potential yet, so once I've used it more, I'll write an update to this article or a follow-up post! Perhaps o3-mini will be released soon, so I'll include my impressions of that as well.

Conclusion: How I use them

• When I want to have a deep, fast-paced consultation with a 'Great Sage,' I use o1 and Gemini 2.0 Experimental Advanced (I'll refer to it as gemini-2.0-exp from here on because it's long).

• When I want to throw a large number of questions at a 'Great Sage' and leave them to process, I use o1-pro.

• For daily consultations and detailed Japanese language tasks (like proofreading articles), a 'Sage' is sufficient. I utilize Claude 3.5 Sonnet (I'll refer to it as sonnet from here on because it's long).

• When I need diagrams or web interface design, I use Sonnet. It creates interfaces nicely in React and draws diagrams well using Mermaid. It makes design and prototyping extremely efficient.

• When I want to see a wide range of opinions, I open multiple windows and send prompts to o1-pro, gemini-2.0-exp, and sonnet, then have someone (usually sonnet) summarize the responses to brush up the ideas.

  • This is especially effective when the path forward isn't clear or when there are many ways to do something.

• For a search engine, Perplexity Pro is the only choice. I sometimes use SearchGPT or DeepResearch, but I rarely rely on them.

  • However, if you already know exactly what you are looking for, existing search engines like Google Search are faster.

• Regarding coding, I often have Sonnet write code while using my own tools in parallel. In those cases, I limit it to things like 'only create small modules.'

  • I think Cline and Recline are worth considering, but I feel existing extensions are a bit lackluster, so I'll probably have to create a fork myself to be satisfied.

  • In that case, I'll just use Sonnet Desktop + MCP while referencing the Cline mechanism.

Now, let's look at each LLM to understand the reasons behind this conclusion. I will not include any discussion of using them via API this time.

OpenAI

gpt-4o and gpt-4o-mini do not have good performance. gpt-4o-mini, in particular, is quite difficult to use. gpt-4o also strongly feels like a 'previous generation AI.' Its instruction-following ability is low.

o1 is extremely smart. It is truly a 'Great Sage.' Honestly, I don't really understand the difference between it and o1-pro. Since o1-pro takes several minutes to return a response, I can only use it by opening many windows and leaving them to process.

o1 makes it easy to ask rough questions. It's easy to then ask deeper questions based on the insights gained from that.

I have judged SearchGPT to be unusable for me, as it is a combination of search-polluted sources and gpt-4o.

Anthropic Claude

3.5-Sonnet is still smart. While it has aspects where it is clearly inferior to o1 and gemini-2.0-exp, its Japanese language handling ability is very strong, and its instruction-following capability is just as good.

Even in the era of o1 and gemini-2.0-exp, it still has its uses. It is still useful for coding assistance, writing Japanese text, or proofreading.

That said, it feels like it is becoming a generation behind, so I would like a new model soon. If a 4 sonnet or something similar were released, it would be strong. I wonder if they will do a thinking/reasoning model as well.

Also, Artifacts is convenient. It casually creates screens using React or diagrams using Mermaid, which saves a lot of time in frontend development and design. This is something you should definitely use.

Gemini

gemini-1.5-pro does not perform well. It is quite difficult to use.

gemini-2.0-experimental-advanced is seriously smart. I feel its intelligence is above sonnet and below o1. On top of that, it is extremely fast. I have only just subscribed, so I haven't fully utilized it yet, and I think this evaluation will continue to be updated, but as of now, the impression is very good. This is worth the subscription.

Maybe my way of using it is completely off-base, but 1.5pro with Deep Research is unusable.

Perplexity

Perplexity Pro has an incredibly exquisite balance between speed and results. Although it is affected by search engine pollution, for some reason, it is relatively better.

It does not show the intense hallucinations seen in Deep Research, the search targets are much better, the output text is good, and above all, it is fast.

Summary

Whether or not you can use the 'Great Sages' (o1 and gemini-2.0-exp-ad) significantly changes a person's intellectual activities. If you don't want to subscribe to ChatGPT Pro, you should subscribe to Gemini Advanced. Gemini Advanced offers a free first month, so you can try it out.

For those who want to perform precise manipulation of Japanese text, use it for coding, or create web screens and diagrams, Claude Professional is recommended. It is inferior to the 'Great Sages,' but you can enjoy conversations with a 'Sage.' It is just that the 'Great Sages' are too strong; the 'Sages' are still plenty smart. At the very least, it is a hundred billion times better than using the 'idiot' models.

For search-related tasks, Perplexity Pro is the only choice. Deep Research will likely improve in the future, but as of now, it is a 'no-go.' Softbank is running a one-year free campaign, so you should try using it.

I would like to write an article about my know-how on utilizing the 'Great Sages' on another occasion later.