DeepSeek showed up in most people’s feeds in January 2025, when a free app from a Hangzhou lab almost nobody in the West had heard of shot to the top of the US App Store and briefly pushed ChatGPT off the number one spot. The attention was partly about the app and mostly about the receipts: a model reportedly trained for a few million dollars that could trade punches with systems costing a hundred times more.
A year and several model releases later, the question is simpler. Is the DeepSeek AI chatbot worth a permanent place in your rotation, or was it a one-month novelty? The honest answer depends on what you ask it to do, and on how you feel about where your prompts end up.
What DeepSeek actually is
DeepSeek is an AI research company spun out of High-Flyer, a Chinese quantitative hedge fund that stockpiled thousands of GPUs before export controls tightened. It ships two things that matter to ordinary users: a free chat app on iOS, Android and the web, and open model weights anyone can download and run.
The models behind the app moved fast. V3 arrived in December 2024 as a 671-billion-parameter mixture-of-experts system that activates only about 37 billion parameters per token, which is how it stays affordable to serve. R1 landed in January 2025 and was the first openly released model to hold its own against OpenAI’s o1 on maths and coding benchmarks. Later versions folded both behaviours into a single model, and that’s roughly what you’re chatting with today.
The $5.6 million headline, with caveats
DeepSeek’s V3 paper put the final training run at around $5.6 million in GPU time. That figure leaves out salaries, data work, failed experiments and the cluster itself, so it isn’t the true cost of building the thing. It’s still a startling number next to the nine-figure budgets Western labs were quoting, and it reset how the industry talked about the price of frontier AI.
The feature you’ll actually notice: thinking mode
Flip on DeepThink in the app and the model writes out a reasoning trace before it answers. Watching it work is oddly absorbing, and occasionally revealing. You can see it reconsider units halfway through, back out of a bad assumption, or catch its own arithmetic slip before committing to a number.
Ask it to debug a 40-line Python function with an off-by-one error buried in a while loop and you get a narrated walkthrough rather than a corrected block dropped in your lap. Ask it the bat-and-ball question and it slows down instead of blurting out the obvious wrong answer. Paste a 60-page contract and ask which clauses shift liability, and it quotes them back with section numbers attached.
That reasoning also exists at small scale. DeepSeek published distilled versions of R1 from 1.5B up to 70B parameters under permissive licences, so a 7B variant runs comfortably on a decent laptop. A local model you can point at confidential files without sending them anywhere is a genuinely different kind of useful.
Where the DeepSeek chatbot earns its keep
- Maths and structured logic. Word problems, unit conversions, probability puzzles. It shows its working, so you can check it.
- Code review and refactors. It reads whole files without complaint and explains why a change matters, not just what to type.
- Long documents. The 128K context window swallows a thesis chapter or a stack of earnings reports in one pass.
- Cost-sensitive automation. Run thousands of calls a day and the API pricing changes whether a project is viable at all.
Where it stumbles
Censorship is the obvious one. Ask about certain dates, certain names, or the status of Taiwan and answers turn vague, or the model pivots to something safer mid-sentence. The inconsistency is arguably worse than a hard block: you can’t tell whether you’re reading a real answer or a deflection unless you already know the subject well.
Beyond politics, it overthinks. A simple question can trigger 30 seconds of internal monologue before a two-line response. It won’t generate images. Web search works but lags on breaking news. And during the January 2025 surge, ‘server is busy’ errors were constant for days.
The privacy question deserves a slow read
DeepSeek’s privacy policy says personal data is stored on servers in China. Several governments responded by banning the app from official devices, and Italy’s regulator blocked it outright before later approving a revised version. If your prompts carry client names, medical details or source code under NDA, think carefully before pasting.
The same caution applies to what you ask of it. Stanford researchers found that mainstream chatbots regularly fail to push back when people bring them serious personal problems, and a model with no memory of your circumstances makes a poor stand-in for a professional. Their work on the dangers of asking AI chatbots for personal advice is worth twenty minutes before you treat any assistant as a counsellor.
Using the API, or running it yourself
The API is where DeepSeek’s reputation was really built. At the time of writing, input tokens cost well under a dollar per million on a cache miss and closer to three cents on a cache hit, with output in the same bracket. Comparable frontier models charge an order of magnitude more. For a support bot or a document pipeline, that gap is often the difference between shipping and shelving.
Open weights mean you can host it yourself too, though that brings hardware headaches. Serving a large model means juggling two very different workloads on the same silicon, a split explained well in this breakdown of why prefill and decode shouldn’t share a GPU. Most teams should start on the hosted API and only self-host when data rules force the issue.
Wire DeepSeek into anything more ambitious and the design question shifts from which model to how many. One model with decent tools covers most workflows; multi-agent systems only earn their overhead when work genuinely splits into parallel specialists.
How it compares with ChatGPT, Claude and Gemini
DeepSeek wins on price, open weights and reasoning per dollar. It loses on polish, image generation, ecosystem and institutional trust. ChatGPT still has the better app. Claude tends to write more carefully and hold a long brief better. Gemini is threaded through Google’s products in a way DeepSeek can’t imitate.
There’s no universal winner, which is the whole point. The best AI chatbot is rarely the one with the biggest launch, and it’s usually the one that fits the shape of your actual work. If you need something narrow instead, like a model tuned for spreadsheets or a context window big enough to swallow a codebase, this roundup of ChatGPT alternatives that beat it at one specific job gets you there faster than testing everything yourself.
A sensible way to slot it into your week
Treat DeepSeek as a second opinion rather than a daily driver. Draft in whichever assistant you already trust, then paste the fiddly parts across and ask DeepSeek to find the flaw. Use thinking mode for anything with numbers in it, because that’s where the extra seconds pay off. Keep confidential material out of the free app unless your employer is comfortable with Chinese data residency. Try the API on one small internal tool before committing a budget.
The larger lesson from DeepSeek’s rise isn’t that a cheap model matched an expensive one. It’s that the gap between top-tier and good-enough reasoning closed faster than anyone’s pricing page could keep up with. That fact alone is worth knowing, whether or not you ever install the app.