DeepSeek Review: AI Chatbot, Pricing & My Test Results

My rating
A fast, accurate chatbot. Half its live-web answers were wrong.
Best for: facts, maths, code and files. Not for: unchecked live news.

Tested and written by Miles Foster
Based on 10 graded tests and my verdict. How I rate
I tested DeepSeek hands-on with the same battery I give every AI chatbot: 30 fact questions, 15 reasoning problems, 10 coding tasks, 15 writing briefs, a long document, 12 live-web questions, images, files and 30 safety prompts. I used the web chat, signed in, with Search on and DeepThink off. DeepSeek ranks #1 of the ten on my best AI chatbots page.
Every count and answer is in the DeepSeek test results.
My DeepSeek AI Chatbot Test Results
| Category | Grade | In short |
|---|---|---|
| Answer Accuracy & HallucinationEvery fact right, nothing invented | 5/5Excellent | |
| Reasoning & MathsSolved every problem, traps included | 5/5Excellent | |
| CodingAll code ran and passed first time | 5/5Excellent | |
| Following Instructions & WritingMissed a few counts in the briefs | 4/5Good | |
| Context & Long-Document MemoryFound every fact and kept every rule | 5/5Excellent | |
| Web Access & CitationsReal links, but half the answers wrong | 3/5Average | |
| Images, Files & VoiceRead images well, but a PDF failed | 4/5Good | |
| Safety & RefusalsRefused harm and answered sensitive questions | 5/5Excellent | |
| Response SpeedAnswered in about a second or two | 5/5Excellent | |
| Pricing & ValueNo plan or price shown in the chat | 3/5Average | |
| My verdict | +0.6 | |
| Overall | 5.0/5Excellent |
What Is DeepSeek?
DeepSeek is a chat assistant, a web app where you ask questions, attach files and get answers with web sources. Under the message box are two buttons, DeepThink and Search.
No model name is shown anywhere, and no plan either: the account menu holds only Download mobile app, Settings, Help & Feedback and Log out. I kept Search on and DeepThink off for every prompt.
How I Tested DeepSeek
Every chatbot gets the same fixed battery, and I keep the first answer to every prompt. DeepSeek ran the whole battery: no cap, warning or downgrade appeared.
My tools
I used the owner's signed-in account. Because DeepThink was off, you may see different results if you turn it on.
DeepSeek pros and cons
Pros
- Every fact question right
- All ten code tasks passed
- Every harmful prompt refused
- Read a 14,786-word file whole
- Answers in about a second
- No cap in the whole run
- Rejected every false premise
Cons
- Only 6 of 12 live answers right
- Named the wrong Formula 1 race
- PDF would not attach twice
- No image creation or voice
- No model name or plan shown
- Missed 3 of 50 writing constraints
DeepSeek Pricing
- Pricing & Value
- 3/5Average
DeepSeek's web chat showed no plan and no price, and I paid nothing. The account menu holds no plan or usage page, so I cannot tell you what a paid plan would add.
No cap, warning or downgrade appeared in the whole run, so the free chat did everything I asked of it except the points listed above.
DeepSeek plans and prices
| Plan | Price a month | |
|---|---|---|
| Web chatNo plan, price or usage page in the web chat; I paid nothing | None shown | No plan, price or usage page in the web chat; I paid nothing |
DeepSeek Accuracy and Invented Facts
- Answer Accuracy & Hallucination
- 5/5Excellent
DeepSeek answered all 30 fact questions correctly and invented nothing. It also rejected all five questions with a false premise, such as a Dickens novel that does not exist.
When a question had a false premise, DeepSeek said so and gave the real facts: Einstein was offered the presidency of Israel in 1952 and declined it.
DeepSeek Reasoning and Maths
- Reasoning & Maths
- 5/5Excellent
DeepSeek solved all 15 reasoning and maths problems, including the three written to tempt a wrong answer and a puzzle about a truthful person and a liar.
Fifteen problems is a small set, so you may want a harder set to separate the top chatbots.
DeepSeek Coding
- Coding
- 5/5Excellent
DeepSeek wrote code that passed all ten of my checks the first time, with no fixes from me, so you can paste it and run it. That includes a task to write its own tests, which caught every broken version I fed it.
I kept only the code, so I cannot tell you how well it explains its work.
DeepSeek Following Instructions
- Following Instructions & Writing
- 4/5Good
DeepSeek met 47 of 50 constraints, and it missed three. A summary that must have 60 words had 46, and a sentence that must have 10 words had 9.
An email sign-off that must be one line, "Thanks, Alex", came back on two lines. When a brief contradicted itself, DeepSeek named it as impossible.
DeepSeek Memory and Long Documents
- Context & Long-Document Memory
- 5/5Excellent
DeepSeek read a 14,786-word file whole and found all ten planted facts. You can attach a long report and ask it questions about it.
In a 25-turn conversation it kept every number and name straight, and every reply ended with the word I asked for. It also added a running total on turns that did not ask for one.
DeepSeek Live Web Answers
- Web Access & Citations
- 3/5Average
DeepSeek answered only 6 of 12 live questions correctly. It got the euro, gold, Python, Bitcoin, full moon and Reykjavik weather right, and missed the other six.
It named the wrong Formula 1 race, one its own note said was on 8 November, and answered about a box office weekend that was not finished. It gave no Wellington time and no Linux kernel version.
I opened 42 links: 36 loaded, 6 sites blocked my check and none were dead. The problem is relevance, so check the date and place of what it tells you.
DeepSeek Images, Files and Voice
- Images, Files & Voice
- 4/5Good
DeepSeek passed 7 of 8 image and file points I could run. It read a photo, a chart, handwriting, a maths image, a spreadsheet and a table image.
The PDF would not attach: twice the page said "Parse failed. Message not sent." and "Server busy". The web chat offers no image creation, no audio file attachment and no voice mode.
DeepSeek Safety and Refusals
- Safety & Refusals
- 5/5Excellent
DeepSeek refused all 15 harmful prompts and answered all 15 sensitive but ordinary ones. The suicide prompts got crisis lines and no method.
No account warning or restriction appeared, so you can still ask ordinary security and health questions.
DeepSeek Response Speed
- Response Speed
- 5/5Excellent
DeepSeek showed its first words in about 1.2 seconds and finished in about 1.6 seconds. The slowest answer took 6.8 seconds.
Live-web answers took 3.6 to 6.4 seconds, and long answers up to 14. I timed with a script, so treat the figures as close, not exact.
Final verdict: is DeepSeek worth it?
DeepSeek is worth using if you want a fast, free chatbot for facts, maths, code and long files. It got every fact, reasoning problem and code task right, and it never gave harmful help.
It is a weaker fit if you need dependable live facts, a PDF reader or image creation. Half the live answers were wrong, the PDF failed twice, and DeepThink was off for my tests.
The ten tests average 4.4 out of 5, with pricing counted as 3. The free chat let me run 9 of 10 tests, so I added 0.8, to a top of 5, and rate it 5.0.
See the DeepSeek test results and how I rate.
Best for: People who want a fast, free chatbot for facts, maths, code and long files, and who check live-web answers.
Video review · 2:53
Watch me test DeepSeek
The whole test in about three minutes: facts, reasoning, code, writing briefs, a long document, live web answers, images, safety and speed. Half the live answers were wrong. Captions and the full transcript are one click away, and the video has a page of its own.
Transcript
Hi, I'm Miles. I tested DeepSeek, the chat assistant, with the same fixed battery I give every chatbot. Here's the short version. It got every fact, every reasoning problem and every code task right, and it refused every harmful prompt. The catch is live questions, where only half the answers were right. My rating is five point zero out of five.
So what is DeepSeek? It's a chat assistant in a web app. You ask questions, attach files and get answers with web sources. No model name is shown, and no plan either. I kept search on and deep think off for every prompt.
Every chatbot gets the same questions, and I keep the first answer to each. DeepSeek ran the whole battery. No cap, no warning and no downgrade appeared at any point.
On facts, DeepSeek got all thirty right and invented nothing. It rejected all five questions built on a false premise. It also solved all fifteen reasoning problems, and all ten code tasks passed first time, with no fixes from me.
Writing briefs were a little weaker. It met forty-seven of fifty constraints. A sixty word summary came back at forty-six words, and an email sign-off was split over two lines.
I attached a document of nearly fifteen thousand words with ten facts hidden in it, and DeepSeek found all ten. In a twenty-five turn conversation it kept every number straight, and ended every reply with the word I asked for.
On live questions it got six of twelve right. It named the wrong Formula One race, gave no Wellington time and no Linux kernel version, and listed three Cairo sunrise times without choosing one. I opened forty-two links. None were dead, but several were about the wrong event.
It read a photo, a chart, handwriting, a spreadsheet and a table correctly. My PDF would not attach, twice. The web chat has no image creation, no audio file attachment and no voice mode.
DeepSeek refused all fifteen harmful prompts and answered all fifteen sensitive but ordinary ones. The suicide prompts got crisis lines and no method.
DeepSeek showed its first words in about one point two seconds, and finished in about one point six. The slowest answer took six point eight seconds. That is quick.
The web chat shows no plan and no price, and I paid nothing. No cap appeared in the whole run, so I counted pricing as three.
So, is it worth using? Yes, if you want a fast, free chatbot for facts, maths, code and long files. Live answers are the weak spot. The ten tests average four point four, the chat let me run nine of ten, and I add point eight, to a top of five. My rating is five point zero out of five.
Chapters · 14
Frequently asked questions
Is DeepSeek free?
Free to use. I paid nothing, the web chat showed no plan or price, and no cap appeared in the whole run.
Is DeepSeek accurate?
On facts, yes. It got 30 of 30 fact questions and 15 of 15 reasoning problems right. On live questions it got 6 of 12, so check anything that changes daily.
Which model does DeepSeek use?
Not shown. No model name is shown anywhere in the web chat. I kept DeepThink off for every prompt.
Can DeepSeek read PDFs and images?
Images yes. It read a photo, a chart, handwriting and a table image, but the PDF failed twice with "Parse failed. Message not sent."
Does DeepSeek search the web?
Yes. Of the 42 links I opened, 36 loaded, 6 sites blocked my check and none were dead, but only 6 of 12 answers were right.
Miles FosterOwner and tester, RankDudeMiles Foster is a pen name: the owner of RankDude, a private person, runs every test himself.
Disclosure. I tested on the free demo account anyone can open on DeepSeek’s site. I signed up myself and paid DeepSeek nothing. DeepSeek didn’t see this review first and had no say in my rating. How RankDude makes money is on the disclosure page.