TL;DR: DeepSeek makes the cheapest strong models in 2026. The API centers on two models: V4-Flash at $0.14 input and $0.28 output per million tokens, and V4-Pro at $0.435/$0.87. Both have a 1M context window and a built-in thinking mode, which is the old R1 reasoning. The open weights are MIT licensed, so you can self-host them for free.
DeepSeek’s whole story is strong models at a very low price. There are two ways to use it. The API runs on two models: V4-Flash is the cheap everyday default, and V4-Pro is the stronger tier for the hardest work. Both include a thinking mode that reasons step by step, which is where the old R1 model now lives. The other way is the open weights, released under the permissive MIT license, which you can download and run on your own hardware for free.
DeepSeek is often 10 to 20 times cheaper than a Western flagship for similar quality, and its weights are MIT licensed, so you can self-host with no restriction. But running it well needs real engineers. This guide helps your team pick between the API and self-hosting, the same way we match engineers to roles.
Key takeaways
- Two API models: V4-Flash ($0.14/$0.28) for everyday work, V4-Pro ($0.435/$0.87) for the hardest tasks.
- Both have a 1M context window, up from 128K in the previous generation.
- Thinking mode is the old R1 reasoning, now built into V4 at a lower price.
- Cache hits cost about $0.0028 per million input tokens, roughly 98% cheaper than a cache miss.
- The open weights are MIT licensed, so you can self-host V3.2 or R1 with no restriction.
The DeepSeek API models at a glance
Here are the two API models, with the model names your engineers will call. The older aliases still work during a transition. Use this table as your reference. Prices are input then output, per one million tokens, at the standard cache-miss rate.
| Model | Context | Price in / out (per 1M) | Best for |
|---|---|---|---|
| DeepSeek-V4-Pro | 1M | $0.435 / $0.87 | Hardest reasoning and coding |
| DeepSeek-V4-Flash | 1M | $0.14 / $0.28 | The cheap everyday default |
| deepseek-chat | 1M | $0.14 / $0.28 | V4-Flash, non-thinking mode |
| deepseek-reasoner | 1M | $0.14 / $0.28 | V4-Flash, thinking mode |
| Cache hit (input) | — | $0.0028 / — | Repeated context, near free |
How do you want to run DeepSeek?
Pick your setup to see the right model and the team you need.
For chatbots, tagging, and routing at scale, V4-Flash is one of the cheapest strong models at $0.14/$0.28 per million tokens, and cache hits are near free. See what building this in-house costs on our developer rate card.
For full data control and zero per-token cost, run the MIT-licensed V3.2 or R1 weights on your own GPUs. This needs strong deployment skills. Hire an AI model deployment engineer who runs vLLM and SGLang in production.
For math, code, and hard logic, turn on thinking mode or step up to V4-Pro. Both reason step by step before answering. Hire AI agent developers who build with them.
The model is the easy part. The hard part is the team that deploys and ships with it. We match you with pre-vetted engineers in about 24 hours. Tell us what you need.
The DeepSeek lineup in plain English
DeepSeek keeps the API simple. There are two models and two modes. V4-Flash is the cheap default. V4-Pro is the stronger tier. Each can run in a fast non-thinking mode or a slower thinking mode. The open weights sit alongside the API for teams that want to self-host. Once you see that shape, the naming stops feeling confusing.
V4-Flash and V4-Pro
V4-Flash is the model most teams should default to. At $0.14 input and $0.28 output per million tokens, with a 1M context window, it handles chat, extraction, tagging, and light coding well. V4-Pro is the stronger tier at $0.435 input and $0.87 output. It is built for the hardest reasoning and the most demanding coding. Even the Pro tier costs a fraction of a Western flagship, which is why cost-sensitive teams favor DeepSeek.
Thinking mode and the R1 story
DeepSeek made its name with R1, a reasoning model that thought step by step. R1 no longer exists as a separate product. Its reasoning now lives inside V4 as a thinking mode you switch on per request. The old deepseek-reasoner alias points to V4-Flash in thinking mode, and deepseek-chat points to the non-thinking mode. Same reasoning quality, lower price, one model.
Open weights and self-hosting
DeepSeek publishes its models as open weights under the MIT license, which is more permissive than most. You can download, modify, and self-host them for any use, including commercial, with no special permission. DeepSeek-V3.2 is a 685B Mixture-of-Experts model, and R1 is 671B with 37B active parameters. Both are freely available on Hugging Face.
Self-hosting gives you full data control and zero per-token cost, which matters for private data and very high volume. The trade is that you run the GPUs and manage the serving stack. That takes real skill. Running a 685B MoE model efficiently is a job for strong AI and machine learning engineers.
Pricing compared across the lineup
DeepSeek is built to be cheap. Output tokens drive most cost, and V4-Flash output is $0.28 per million, with V4-Pro at $0.87. Both are far below a Western flagship. The chart shows input and output price for the two API models.
Two extra levers make it cheaper still. Cache hits charge repeated context at about $0.0028 per million input tokens, roughly 98% off a cache miss, which is huge for agents that reuse a big system prompt. DeepSeek also runs off-peak discounts at quieter hours. Teams that use caching and off-peak batching well cut their bill far below the sticker rate. This is the kind of work that separates a strong AI engineer from an average one.
Context windows and thinking mode
Both V4 models offer a 1M token context window, a big jump from the 128K to 164K of the previous generation. That is roughly 750,000 words, enough for full-codebase reasoning and long multi-document work in a single request. Max output on both is 384K tokens, so even long generations fit.
Thinking mode is the main quality lever. Turn it on for hard math, code, and logic, where step-by-step reasoning pays off. Leave it off for simple chat and tagging, where speed and cost matter more. Thinking mode uses more output tokens, so it costs more per call. An engineer who switches modes on purpose runs DeepSeek far cheaper at the same quality. You can read more in our breakdown of Claude versus ChatGPT for coding.
Which model should you use
Start with V4-Flash in non-thinking mode. It is the cheap default for most tasks. Turn on thinking mode when a task needs reasoning. Step up to V4-Pro for the hardest work. Self-host the open weights when you need data control or run very high volume. The table below maps common jobs to a model.
| Job | Recommended model | Why |
|---|---|---|
| Support chatbot, FAQ | V4-Flash (non-thinking) | One of the cheapest strong models |
| Data tagging, extraction | V4-Flash (non-thinking) | Fast, cheap, 1M context |
| Math, logic, planning | V4-Flash (thinking) | R1-grade reasoning, low price |
| Hardest reasoning and coding | V4-Pro | Stronger tier, still cheap |
| Private data, high volume | V3.2 or R1 (self-host) | MIT license, $0 per token |
| Agents with big prompts | V4-Flash + caching | Cache hits are near free |
A simple rule helps. Prototype on V4-Flash, measure your real volume, then decide if thinking mode, V4-Pro, or self-hosting pays off. We worked with a team that ran everything on a Western flagship. Moving routine work to DeepSeek V4-Flash and keeping only the hardest calls on a premium model cut their bill by more than half with no drop in quality.
Older DeepSeek models
DeepSeek ships fast, and older versions are being folded into V4. If your code calls an older model or alias, plan to move it. The deepseek-chat and deepseek-reasoner names are set to change, so pin the V4 model names when you can. The table shows the main older models and what to use instead.
| Older model | Status | Move to |
|---|---|---|
| DeepSeek-R1 | Folded into V4 | V4 thinking mode |
| DeepSeek-V3, V3.1, V3.2 | Superseded on API | V4-Flash or V4-Pro |
| deepseek-chat, deepseek-reasoner | Aliases, changing | V4-Flash by name |
| DeepSeek-V2, Coder | Legacy | V4 or self-hosted V3.2 |
A move is usually small. On the API you change the model name. For open weights you download the new checkpoint and update your serving config. The DeepSeek API docs and the model cards on Hugging Face show every current and older version in one place, so check them before you migrate.
What this means for hiring
Knowing the lineup is step one. Getting value from it is the harder step. With DeepSeek the biggest levers are thinking mode, caching, off-peak batching, and the choice to self-host. The teams that win are the ones who use these well, not the ones on the biggest model. That is engineering, not just a model choice.
Engineers who can deploy and optimize open models are scarce and costly in the United States. A senior AI engineer there runs $12,000 to $18,000 per month all in. The same skill from a pre-vetted engineer in Vietnam or the Philippines often costs $3,000 to $6,000 per month. The work is the same. The bill is not. Many of our clients build their whole DeepSeek stack with an AI model deployment engineer from our network.
Compare other model families
Build your DeepSeek stack with the right team
DeepSeek gives you strong reasoning at the lowest price, plus MIT open weights you can self-host. The model is the easy part. The team that deploys and ships with it is what creates value. We match you with pre-vetted senior engineers across Asia in about 24 hours, with no upfront cost and payroll handled.
