DeepSeek's line-up is small compared with OpenAI's, which makes it easier to learn. There is a chat family that gets a new version number every few months, a reasoner family that started with R1, and a handful of variants — Flash, Pro, Terminus, distilled models — that sit alongside. This lesson maps them and shows what each costs on the official API and on third-party hosts.
The chat line: V3 to V4
The chat models are what you get by default in the app and as deepseek-chat on the API. DeepSeek V3 was the breakthrough release; V3.1 and V3.2 refined it, and V3.1 Terminus is a stability-focused checkpoint. The V4 family splits into V4 Flash, a small, very cheap model for high-volume work, and V4 Pro, the larger flagship for harder tasks.
At the time of writing our catalogue shows V4 Flash around $0.07–$0.09 per 1M input and roughly $0.18 per 1M output tokens depending on host, V3.2 at about $0.27 in and $0.40 out, and V4 Pro between roughly $0.58 and $1.01 in and $1.74 to $2.02 out depending on checkpoint and host. Those are all far below the closed-model flagships you see on /models.
The reasoner line: R1
DeepSeek R1 is the model that made reasoning models a mainstream topic. It is served as deepseek-reasoner on the official API, and the R1 0528 checkpoint is what most third-party hosts carry. In our catalogue R1 sits at roughly $0.50–$0.70 per 1M input and $2.15–$2.50 per 1M output tokens at the time of writing — still cheap in absolute terms, but note that output is five to six times the price of V3.2's output, and a reasoner produces a lot more of it.
| Model | API name | Best for | Watch out for |
|---|---|---|---|
| DeepSeek V4 Flash | deepseek-chat (or host-specific id) | Bulk summarising, classification, extraction, simple chat | Weaker on hard reasoning; check quality on a sample first |
| DeepSeek V3.2 | deepseek-chat | General writing, coding, everyday assistant work | Superseded by V4 on some hosts; prices shift often |
| DeepSeek V4 Pro | deepseek-chat (host id varies) | Complex coding, long documents, agentic tasks | Roughly 2–4x the price of V3.2; confirm you need it |
| DeepSeek R1 / R1 0528 | deepseek-reasoner | Maths, logic puzzles, tricky debugging, planning | Long hidden reasoning inflates output tokens and latency |
| R1 Distill Llama 70B | host-specific id | Cheap reasoning on hosts that carry it | Not an official DeepSeek endpoint; 8K context on some hosts |
Context windows
Context window is how much text a model can hold at once, counting your prompt, any documents you paste, and the answer. Our catalogue lists the V3 line at about 164K tokens, R1 checkpoints between 64K and 164K, and the V4 family at roughly 1M tokens or more. One million tokens is around 750,000 words, enough for a full codebase or a shelf of reports.
Worked example: one task, three models
Suppose you need to classify 1,000 support tickets. Each request is about 2,000 input tokens (instructions plus ticket) and 500 output tokens (a structured label and a one-line reason). That is 2M input and 0.5M output tokens in total. Using the catalogue prices at the time of writing:
- V3.2: 2 × $0.27 + 0.5 × $0.40 ≈ $0.74 for all 1,000 tickets.
- V4 Pro 0813: 2 × $0.58 + 0.5 × $1.74 ≈ $2.03.
- R1 0528: 2 × $0.50 + 0.5 × $2.15 ≈ $2.08 — and in practice more, because the reasoner will spend extra output tokens thinking about each ticket.
None of those numbers is scary, which is the point of DeepSeek. But the pattern is the one you will see everywhere in this course: a chat model is usually the right default, and you escalate to Pro or a reasoner only for the subset of requests that fail on the cheap model. Run your own numbers in the calculator.
Knowledge check
You need to summarise 50,000 short product reviews into one-line summaries. Which model should you try first?