Grok, DeepSeek and Qwen get grouped together as the cheap alternatives to ChatGPT, Claude and Gemini. On price that grouping holds. On almost everything else it falls apart, and the differences are the kind you only find by reading the pricing pages and the terms rather than the headlines.
Three things separate them: whether you can find out what you are buying, what happens to what you type, and whether you can walk away with the model.
Only two of the three publish a number
DeepSeek and Qwen both publish per-token rates you can put in a spreadsheet. Here are the comparable fast models, per million tokens:
| Input | Output | |
|---|---|---|
| deepseek-flash, off-peak | $0.15 | $0.60 |
| deepseek-flash, peak | $0.30 | $1.20 |
| qwen3.8-flash | $0.15 | $0.47 |
| deepseek-v4-pro, off-peak | $0.66 | $1.98 |
| deepseek-v4-pro, peak | $1.32 | $3.96 |
| qwen3.8-max | $2.00 | $6.00 |
Grok publishes nothing comparable. Its consumer plans run $8 for Plus, $30 for Premium, $100 for Super and $200 for Ultra, and the usage each one buys is described only as a multiple: “4x more usage compared to Plus”, then 15x, then 50x. Plus itself is never quantified anywhere on the page. A multiplier with no base is not a number.
It is also a shared pool rather than a per-product allowance: “Every tier includes a pool of AI usage shared across Grok and Cursor”, refreshing weekly. API access comes with the paid plans but is “billed separately at standard model rates and doesn’t draw from your plan’s usage pool”, so a subscription does not cover programmatic use.
DeepSeek charges by the clock
This is the most unusual pricing structure in the group, and the easiest to miss. DeepSeek bills different rates at different times of day. In its own words: “Off-peak rates are half of the peak rates. Peak hours are 01:00 – 04:00 and 06:00 – 10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and Chinese public holidays in full.”
Translated into local time, peak is 05:00–08:00 and 10:00–14:00 in the Gulf, and 06:00–09:00 and 11:00–15:00 in Pakistan. Weekends are off-peak in full, everywhere.
For an interactive assistant this is noise. For batch work — overnight document processing, scheduled summarisation, anything you control the timing of — it is a straight fifty per cent saving for moving a cron job by a few hours. Very few vendors let you do that.
Caching changes the picture again. A cache hit on input drops from $0.15 to $0.003 per million tokens off-peak. If your workload sends the same long prompt repeatedly, that is where the real saving sits, not in the headline rate.
Qwen’s two catches
Qwen’s rates are flat across the day, which is simpler. But the pricing is tiered by request size, and the tier applies to the entire request. Alibaba’s documentation gives the example plainly: “If a request contains 100K input tokens, it falls into the second tier (32K < 100K ≤ 128K), and all tokens are billed at the unit price of the second tier.”
Not the tokens above the threshold. All of them. A request that crosses a tier boundary by one token reprices every token in it. If your prompts sit near a boundary, trimming them is worth more than it looks.
The second catch is geographic. The free quota that appears next to most Qwen models carries this note: “The following models offer a free quota only in Singapore. No free quota is available in other regions.” Where it does apply it is one million tokens, valid for 90 days. If you are evaluating Qwen from anywhere else, there is no free trial — you are paying from the first call.
Worth noting too that some Qwen-Plus rates currently display as “List price … (Limited-time 20% off)”. A promotional price is not a price. Budget on the list figure.
What happens to what you type
DeepSeek is direct about collecting it. Its privacy policy, last updated 10 February 2026, says it may collect “your text input, voice input, prompt, uploaded files, photos, feedback, chat history, or other content that you provide to our model and Services”, and lists among its uses: “To improve and develop the Services and to train and improve our technology, such as our machine learning models and algorithms.”
It then does something several larger competitors do not: it grants an explicit opt-out. The policy lists “the right to opt-out of using your Personal Data for training our models or optimizing our technologies.” It also states flatly that it does not engage in targeted advertising, sell personal data, or profile users. The data controller is Hangzhou DeepSeek Artificial Intelligence Co., Ltd., registered in China — a fact that will matter to some readers and not others, but one you should know before deciding.
Grok’s terms take a different shape. You keep ownership — “you retain your ownership rights to the User Content” — but you grant the company a licence over your inputs that is “irrevocable, perpetual, transferable, sublicensable, royalty-free, and worldwide”, for purposes including to “improve our products and the Service … developing new products or features”. The same rights are extended to “a person’s image, likeness, voice, or other similar attributes”. There is no opt-out in that clause.
One clause in Grok’s terms is worth quoting exactly because it is genuinely unclear: “When using Output or SpaceXSI’s name, logos, trademarks, or other brand elements, you are required to obtain our permission and attribute your generation of the Output to the Service, as detailed in our Brand Guidelines.” Read one way, permission and attribution attach only to the brand elements. Read another, they attach to Output itself. We are not going to resolve it for you; if you are publishing commercially off the back of Grok output, it is a sentence to put in front of a lawyer.
We did not read Qwen’s consumer chat terms for this piece, so we are not making a claim about them. For the hosted API the relevant documents are Alibaba Cloud’s, which are a separate matter from the chat app.
The difference that outlasts the price
Prices change. Licences are harder to take back, and this is where the three genuinely part company.
Qwen publishes open weights under the plain Apache License 2.0 — the README states it without qualification: “All our open-weight models are licensed under Apache 2.0.” The range runs from 0.6B up to a 235B mixture-of-experts model, so there is something that fits most hardware. DeepSeek’s R1 and V3 repositories carry the MIT License. Both are permissive, standard, and come with no territorial carve-out or user-count trigger.
Grok has no downloadable weights. Whatever you build on it stays on their service, under their terms, at their prices.
That is the whole argument, really. With Qwen or DeepSeek, a price rise or a terms change is an annoyance: the weights you already have keep working, and the privacy question stops being a question once the model runs on your own hardware. With Grok, it is a migration.
Which one suits what
- Scheduled or batch work — DeepSeek, run off-peak, with caching if your prompts repeat. The time-of-day discount is real money and nobody else offers it.
- Long prompts at a predictable price — Qwen, as long as you watch the tier boundaries. Flat rates and no clock-watching.
- Anything you want to keep running whatever happens — either of the open-weight pair, self-hosted. Qwen’s smaller sizes make this the more realistic of the two on ordinary hardware.
- Grok — the honest answer is that you cannot budget for it from published information. If you need a committed number before you commit, this is not the one.
What we checked
DeepSeek’s API pricing page and privacy policy, and Alibaba Cloud’s Model Studio pricing documentation, were read on 4 November 2026. xAI’s pricing page and consumer terms of service were read on 3 November 2026. The Qwen3 and DeepSeek-R1 repositories and their licence files were read on 4 November 2026. Everything in quotation marks is taken verbatim from those documents. Prices are in US dollars, per million tokens, for international deployment, and exclude tax. We have not benchmarked any of these models, and nothing here is a claim about output quality.
For the same comparison across the larger assistants, see ChatGPT vs Claude vs Gemini: the same price, three different sets of terms. For what “open source” survives contact with the licence file, see Four self-hosted AI tools called “open source”. Only one of them is.