DeepSeek API pricing starts at $0.14 per 1M input tokens on DeepSeek V4-Flash — around
a sixtieth of what frontier models charge. That gap, not the parameter counts, is why
DeepSeek matters. Every model in the line is also released under an open licence, so the API
price is a convenience charge rather than the only way to run them.
DeepSeek API pricing: every model, per 1M tokens
| Model | Input $/1M | Output $/1M | Blended $/1M | Context |
|---|---|---|---|---|
| DeepSeek V4-Flash open weights | $0.140 | $0.280 | $0.168 | 1M |
| DeepSeek V4-Pro open weights | $0.435 | $0.870 | $0.522 | 1M |
| DeepSeek R1 Distill Llama 70B open weights | $0.800 | $0.800 | $0.800 | 128K |
| DeepSeek R1 open weights | $0.500 | $2.15 | $0.830 | 128K |
Blended is the effective rate at a 4:1 input-to-output mix, which is what a typical chat or retrieval workload actually produces. It is the number to compare across vendors — a headline input price hides how much the output side costs.
What DeepSeek costs per month
| Workload | Tokens / month | DeepSeek V4-Flash cheapest | DeepSeek R1 most capable tier |
|---|---|---|---|
| Side project | 1M in / 0.25M out | $0.21 | $1.04 |
| Small team | 20M in / 5M out | $4.20 | $21 |
| Production | 200M in / 50M out | $42 | $208 |
Model your own volumes in the AI API cost calculator.
How DeepSeek compares to other providers
Each provider's cheapest priced model, so the comparison is like for like on entry cost.
| Provider | Cheapest model | Blended $/1M |
|---|---|---|
| Mistral AI | Mistral 7B | $0.0220 |
| Meta | Llama 3.1 8B | $0.0220 |
| Alibaba | Qwen3 8B | $0.0600 |
| Gemma 3 4B | $0.0600 | |
| Microsoft | Phi-4 | $0.0840 |
| DeepSeek this page | DeepSeek V4-Flash | $0.168 |
| Moonshot AI | Kimi K2.7 Code | $0.980 |
| Anthropic | Claude Haiku 4.5 | $1.80 |
| Zhipu AI | GLM 5.2 | $2.00 |
| xAI | Grok 4 | $5.40 |
| OpenAI | GPT-5.6 Sol | $10.00 |
Cheapest is not the same as best value — check capability alongside price on the LLM leaderboard, or browse every model in the AI models database.
Which DeepSeek model should you use?
V4-Flash is the volume workhorse: 284B total parameters with roughly 13B
active per token, a 1M-token context, and a blended rate near $0.17 per 1M against a frontier
model’s $10. For bulk classification, document processing, first-pass summarisation, synthetic
data generation and the retrieval layer of a RAG pipeline, it is close to unmatched on
capability per dollar. A 1M context at that price does not exist elsewhere.
V4-Pro is the open flagship — a 1.6-trillion-parameter mixture-of-experts
activating around 49B parameters per token, with a 1M-token context. At roughly $0.52 blended it
runs about a twentieth of frontier pricing while holding its own on general reasoning.
R1 is the landmark open reasoning model, a 671B mixture-of-experts with 37B
active, delivering chain-of-thought reasoning at a fraction of proprietary cost. Its 128K
context is narrower than the V4 line’s 1M.
R1 Distill Llama 70B exists for one reason: hardware. Full R1 needs roughly
400 GB of VRAM at 4-bit; the distill needs about 40 GB, which moves it from “rent a server” to
“buy a workstation”. It also bills $0.80 for both input and output, so generation-heavy
workloads cost no more than prompt-heavy ones — unusually simple to budget.
The switching cost is lower than you think
The detail that makes DeepSeek a genuine option rather than a curiosity is that the V4-Pro
API speaks both OpenAI and Anthropic request formats. Migrating is usually a base-URL and key
change rather than a rewrite. The integration friction that normally protects incumbent
providers largely disappears, which is what makes a 20x cost difference actionable instead of
theoretical.
The sensible pattern is not wholesale migration but routing: send the high-volume,
low-stakes majority of traffic to V4-Flash and keep a frontier model for the requests that
genuinely need it. Most applications discover that the split is heavily weighted toward the
cheap side.
Do the open weights change the maths?
Every DeepSeek model here is openly downloadable, which raises the obvious question. For
most teams the honest answer is that the licence buys auditability and provider portability
rather than a realistic on-premises deployment.
The hardware requirements are the reason. V4-Pro needs roughly 800 GB of VRAM at 4-bit — a
data-centre exercise of eight H100 80GBs or more. R1 needs about 400 GB. V4-Flash needs about
140 GB, or two H100 80GBs. Only the R1 distill, at 40 GB, fits hardware a small team would
actually buy.
Set against API pricing this low, self-hosting has to clear a high bar: it only pays once
your volume is high enough to keep that hardware continuously busy, and the break-even point
moves every time DeepSeek cuts prices. Work it out for your own token volume in the
self-hosting vs API calculator
before committing to hardware. What the open weights do guarantee is that you can never be
locked in, price-gouged, or stranded by a deprecation — which is worth something even if you
never download them.
Frequently asked questions
How much does the DeepSeek API cost?
DeepSeek API pricing runs from $0.168 per 1M blended tokens on DeepSeek V4-Flash up to $0.830 on DeepSeek R1. Blended assumes a 4:1 input-to-output mix. Input and output are billed separately, and output is always the more expensive side.
What is the cheapest DeepSeek model?
DeepSeek V4-Flash, at $0.140 per 1M input tokens and $0.280 per 1M output. A small-team workload of 20M input and 5M output tokens a month costs about $4.20 on it.
How much does DeepSeek cost per month?
At 20M input and 5M output tokens a month, DeepSeek V4-Flash costs about $4.20 and DeepSeek R1 about $21. A side project at 1M/0.25M costs a small fraction of that. Use the AI API cost calculator for your own volumes.
Is DeepSeek cheaper than Mistral AI?
On entry-level pricing, DeepSeek starts at $0.168 per 1M blended and Mistral AI starts at $0.0220 on Mistral 7B. Mistral AI is the cheaper entry point, though capability differs — compare both on the leaderboard before switching.
Why does DeepSeek charge more for output tokens than input?
Output tokens are generated one at a time and cannot be batched the way a prompt can, so they cost more to serve. This is why prompt-heavy workloads such as retrieval and classification are far cheaper to run than generation-heavy ones, and why the blended rate matters more than the headline input price.
Prices are the published list rates for each model's primary API and are reviewed as providers change them. Volume, batch and cached-input discounts are not included. Last reviewed August 2026.
