iTranslated by AI
AI Tool Usage and Impressions: Late 2025 Edition
I'm Beagle 🐶
Previously, I wrote an article like this.
Believe it or not, that was an article from 2025.
It felt like I wrote it much longer ago, but that's no surprise. The AI tool environment has been completely changing on a monthly basis.
In this rapidly changing AI landscape, I'll record my usage status as of the end of 2025 in this article.
When I look back later, I'm sure I'll feel nostalgic thinking, "So this is what the end of 2025 was like..."
Also, I would be very happy if my AI tool usage and impressions could be of some reference to someone.
Where It All Began
Specific Use Cases
Consultation, Brainstorming, Technical Explanations, and Daily Questions
In most situations, I ask ChatGPT's GPT-5.2 Thinking.
While I sometimes use Grok for a second opinion or request Claude Code to perform research, GPT-5.2 remains my unwavering core.
Summarization + Knowledge Accumulation
I think the summarization performance of all models has become very high.
Prioritizing ease of use, I have Claude Code's Opus 4.5 create and categorize md files.
Long texts and videos are put into NotebookLM.
Text Generation
For articles and documents, to be honest, I think it's faster and easier to convey intent if I write them myself.
However, in situations where a certain level of formatting or consideration is required, I have various LLMs generate content, compare them, and adopt the one that looks best.
Image Generation
Nano Banana Pro is overwhelmingly strong in every aspect. It's the only choice.
Coding
I've tried many different tools, but I've settled on Claude Code (+ Windsurf).
Claude Code's Opus 4.5 performance is well-balanced and excellent. The agent capabilities of Claude Code itself are high, and its meticulous attention to detail for engineers is quite appealing.
Windsurf offers free code completion, and the SWE-1-mini model it uses for completion seems excellent.
Since my style absolutely requires a code completion feature, I'm grateful for Windsurf.
Specific Subscriptions and Service Impressions
OpenAI
Plan: ChatGPT Plus ($20 / month)
I feel that this year was the era of ChatGPT from beginning to end.
While I had some trouble with GPT-5 because it spoke strange Japanese, the latest model, GPT-5.2, is truly excellent and seems to be the finalized form of a chat model at this point.
I use ChatGPT's GPT-5.2 Thinking for a vast range of tasks, including consultations, brainstorming, questions, technical discussions, and daily doubts.
OpenAI also offers other features like the coding agent Codex, the image generator GPT Image, and the video generator Sora.
Subscribing to Plus gives you preferential access to these functions, making it extremely cost-effective.
I bring in Codex as a secondary coding agent for specific tasks. I feel its performance in reviews and bug investigations is comparable to, or even better than, Claude Code.
Anthropic
Plan: Claude MAX ($100 / month)
While $100 a month might seem insane, thanks to Claude Code, I've achieved what feels like a nearly 10-fold increase in productivity.
Specifically, I was able to complete the release of an OSS project in just two full days, which I had originally estimated would take a month.
It will likely continue to serve as my main weapon for a while in 2026.
I'll consider a higher plan if I run out of capacity, but the current plan is working fine for now.
On the other hand, Opus as a chat model is more mediocre than I expected, and since it shares usage limits with Claude Code, I don't use it very often.
Plan: AI Pro ($20 / month)
Considering the preferential treatment for Nano Banana Pro and NotebookLM, as well as the storage fees, the cost-performance is good.
I don't use Gemini much, but it's not that bad.
X / xAI
Plan: X Premium Annual (6,426 yen / year)
I joined during a discount period. At less than 600 yen per month, it's very cost-effective, offering relaxed usage limits for Grok, access to X Pro, and the verification mark.
Grok's performance has seemed subpar lately, but I expect they are planning diligent upgrades, so I'm looking forward to that.
Windsurf
Plan: Free
As mentioned earlier, the code completion suits my style. If you subscribe to a paid plan, you can also use it in a way similar to Cursor.
Subscriptions I Previously Joined
GitHub Copilot
For a while, after switching from Roo Code, it was my main tool for coding, but I gradually started using other tools.
I'm happy to be able to use various models. It also includes code completion features, and the cost-performance ratio isn't bad.
Cursor
I also used this as my mainstay for a while.
I feel that Cursor has a very high ability to draw out the potential of the models.
So, Which One is Recommended?
If you have the budget, I recommend subscribing to all the $20 plans from the big three model providers (OpenAI, Anthropic, and Google).
When talking about AI, I feel that subscribing to these three $20 plans is the starting point.
For those starting with AI now, I recommend trying at least one of these three.
If only one, I guess it would still be ChatGPT?
I think the performance has reached a point where you can say "any of them will do" to some extent.
If you are mainly focused on coding and utilize an IDE, Cursor is also recommended.
Summary
2025 was (also?) a year of upheaval. My usage style shifted from "one for everything" to "using different tools for different purposes," evolving toward a distributed approach across the three major providers: GPT-5.2 for consultation, Nano Banana Pro for images, and Claude Code for coding. An expenditure of about $140 per month is by no means cheap, but I feel it is an investment well worth it considering the significant improvement in productivity. I expect to continue with this approach next year as well.
Regarding coding agents, I have high expectations for the evolution of Claude Code in 2026. Even now, the benefits to productivity are immeasurable, and I believe it will become an even more fantastic tool as the models evolve and agent capabilities are enhanced.
On the other hand, with the potential rise of OpenAI's Codex and autonomous coding agents, it's quite possible that the competitive landscape will shift into something entirely different sometime next year. I look forward to continuing to learn about AI while enjoying these turbulent times next year as well, so thank you for your continued support.
Discussion