[Honest Review] What I Really Discovered After One Month of Using DeepSeek
機械翻訳 / Machine-translated

機械翻訳 / Machine-translated
@aifriends
AI Friends(https://aifriends.jp)のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
What you'll learn from this article
I started using DeepSeek after hearing all the buzz around "R1," the reasoning-focused model released in January 2025. The pitch — performance on par with paid ChatGPT, completely free — was compelling enough to make me sign up and give it a shot. I was a little wary at first, given that it comes from China, but the fact that it's open-source (meaning anyone can inspect the underlying code) gave me a sense of credibility.
The API costs (the fees charged when accessing the model programmatically) were also a massive draw — reportedly less than one-twenty-fifth of what competitors charge. For someone like me who uses AI heavily for work, keeping running costs down is a big deal. When V4 Pro, the latest version, launched in April 2026 with yet another performance jump, I decided to dive in and use it seriously.
The first thing that surprised me in that opening week was the inference speed. Even when I asked it to work through complex math problems or generate programming code, responses came back almost instantly. For math in particular, it walks through its own thinking process step by step, making it easy to follow exactly how it arrived at the answer.
On the other hand, the naturalness of its Japanese felt slightly off at times. Grammatically it was fine, but it occasionally produced awkward phrasing, and my overall impression was that its Japanese-language accuracy lagged a step behind ChatGPT and Claude. I was also uneasy about security during those first few days, so I was careful never to enter any sensitive information. When a tool is free, questions about how your data is handled tend to linger.
1. A high-performance AI that's completely free
The biggest appeal is the total absence of a monthly fee. ChatGPT Plus costs $20 per month (around ¥3,000), whereas DeepSeek gives you an equivalent level of reasoning AI for ¥0, with no usage cap. For students or anyone who just wants to experiment without committing, this feels like the best option out there.
2. Exceptional at logical reasoning tasks
For tough math problems and algorithm design in programming, it outperformed other AI tools in accuracy. It has earned high marks on benchmarks (standardized performance tests) as well — for example, achieving a 79.8% correct-answer rate on International Mathematical Olympiad problems. It proved genuinely reliable when I needed to work through technical challenges at work.
3. Open-source and cleared for commercial use
The model's internals are publicly available, and you can download and run it locally on your own computer. It's distributed under the MIT License (a set of rules that permits commercial use), which means businesses can adopt it with confidence — a significant advantage.
1. Japanese-language quality leaves something to be desired
When generating Japanese text, unnatural expressions and mistranslations were noticeable. Especially in situations that call for business documents or polished, courteous language, I found ChatGPT and Claude to be clearly superior. A Japanese-specialized model was released in January 2025, but there's still room for improvement.
2. Security and privacy concerns
Being a Chinese-made AI, there are legitimate questions about how data is handled. A large-scale cyberattack incident occurred in January 2025, and an increasing number of major Japanese companies have banned its internal use. It is simply not safe to enter personal information or trade secrets under any circumstances.
3. Meaningful barriers to enterprise use
In industries such as finance, healthcare, and legal services, security policies (internal rules governing data safety) often restrict or prohibit the use of DeepSeek altogether. Even if there's no issue with personal use, using it at work will frequently require sign-off from a manager or the IT department.
Compared with ChatGPT (GPT-4), DeepSeek has the edge in reasoning depth, but ChatGPT wins on conversational naturalness and Japanese-language quality. For instance, DeepSeek's logical analysis was genuinely useful when hunting down bugs in code, but when it came to drafting customer-facing emails, ChatGPT was the more comfortable choice.
On the API pricing front, DeepSeek wins decisively. Tasks that might cost several thousand yen a month through the ChatGPT API sometimes cost only a few dozen yen with DeepSeek. That said, while the free tier is said to have no usage limits, I did experience sudden, unexplained outages — an instability worth keeping in mind.
Recommended for
Not recommended for
DeepSeek is far from perfect, but the fact that you can get this much out of it for free is genuinely remarkable. Figuring out whether it suits your particular needs and pairing it with other AI tools where appropriate feels like the best strategy for now.
This article is a cross-post from AI Friends.