DeepSeek
Open-weight reasoning models with a free chat interface.
What it is
DeepSeek is the strongest free option for reasoning-heavy work — maths, code, step-by-step analysis. The weights are published, so you can also self-host, which is the real differentiator if data cannot leave your infrastructure. Its interface is plainer than the paid leaders, and that is the trade.
Pricing updated 2026-11-10 Official pricing
DeepSeek pricing explained
There is essentially one free product and a metered API. The chat interface costs nothing and includes the reasoning model, which is unusual — most vendors put reasoning behind a subscription. The API is billed per token at rates that make it viable to call from a real product rather than only from a notebook. If you need a guaranteed service level, a team workspace or an SLA, this is not the vendor to pick; the trade for the low price is that those things are absent, and peak-hour latency is visible on the free chat. Confirm current pricing on the official API page before building a budget around it.
Free-tier limits are taken from the official pricing page on the update date above. They change often — confirm before you commit to an annual plan.
DeepSeek key features
Reasoning with visible steps
The thinking trace is shown before the answer, so you can see where it took a wrong turn instead of only seeing the wrong conclusion.
Open weights — self-hostable
The model files are published. If data cannot leave your network, you can run the same model on your own hardware instead of arguing about a vendor contract.
Coding assistance
Handles refactors and debugging across files with the reasoning trace on, which is where the step-by-step mode earns its keep.
File upload
Attach documents and ask questions about them. Basic compared with the paid leaders, but enough for reading a spec or a paper.
Web search
Can pull current results when a question needs them, though the citations are thinner than what the search-native assistants produce.
Pay-as-you-go API
The same models behind a per-token price, which is the cheapest way to put a reasoning model inside your own product.
DeepSeek: how it works
Mixture of experts
Only a fraction of the network activates per token, so a large model runs at a small model cost. That is the reason a free tier this capable is economically possible at all.
Attention designed for long context on a budget
A compressed attention scheme keeps the KV cache small enough to serve long prompts without the memory bill of a dense model. The trade shows up as a slightly flatter recall on very long inputs.
Reasoning trained by reinforcement learning
The reasoning behaviour comes from RL on verifiable tasks rather than from imitation alone, which is why maths and code improve sharply while conversational polish does not.
Weights published under a permissive licence
Downloadable weights change the deployment story: quantise it, run it air-gapped, fine-tune it. It also means anyone can audit the claims about a model that a hosted API would hide.
DeepSeek core strengths
The reasoning model is the free model
No cut-down version behind a paywall. For maths, code and structured analysis, the free tier is the strongest of the group.
Open weights allow self-hosting and auditing
A real option when regulation, procurement or paranoia rules out a hosted API. Few competitors in this list offer it at all.
Strong at code and step-by-step reasoning
The visible trace makes it useful for debugging logic, not just for producing text that looks plausible.
Cheap at the API level
Per-token pricing well below the US labs means a side project can afford to call a reasoning model in production.
DeepSeek compared with alternatives
| Tool | DeepSeek | ChatGPT | Claude | Qwen Chat |
|---|---|---|---|---|
| Best for | Reasoning-heavy work on a budget, or self-hosting for data control | Mixed daily work: writing, search, files, images | Long documents, and writing that survives several revisions | Bilingual English and Chinese work, with images and documents |
| Free tier | Free chat, no paid tier required | Usable but capped; newest models restricted | Daily message cap; uploads limited | Free with generous limits |
| Input and output | Text only; no image generation | Text, image and voice in; text, image and voice out | Text and image in; text out, no image generation | Text, image and document in; text out |
| Ecosystem and integration | Open weights plus a low per-token API | Custom GPTs, documented API, first-party desktop and mobile | Projects, Artifacts, API, MCP; smaller plugin ecosystem | Open-weight family plus API |
| Paid entry price | Free; API billed per token | $20 / month | $20 / month | Free; API billed per token |
Written from the official product pages listed under Sources, not copied from another directory. Prices and limits reflect the update date above; features ship faster than reviews, so treat the list as a starting point and confirm on the official site.
Pros and cons
Strengths
- Free tier includes the reasoning model, not a cut-down one
- Open weights allow self-hosting and auditing
- Very strong at code and structured reasoning
Limitations
- Interface and integrations are basic
- Slower at peak times than paid tiers
- Fewer multimodal features than the leaders
DeepSeek use cases
Maths and quantitative checking
Paste a derivation or a spreadsheet formula and ask where the step is wrong. The visible reasoning makes the error findable instead of just producing a different number.
Refactoring with constraints
Give it the module, the constraint and the target shape; the trace shows whether it understood the constraint before it started editing.
Air-gapped deployment
Download the weights and run them on internal hardware when the data cannot cross a network boundary. This is the reason many teams shortlist it at all.
Cheap batch processing
Classifying or summarising thousands of records through the API at a price that does not require a business case.
Learning a topic by interrogation
Ask it to reason through a problem and then read the trace to see which assumption it made — a faster way to find your own gap than reading an explanation.
