DeepSeek API Pricing Sheet & Token Rates 2026

DeepSeek disrupted global AI economics by pricing frontier-grade reasoning (R1) and general knowledge (V3) models at up to 96% lower cost than proprietary US frontier labs, backed by native server-side prompt caching.

DeepSeek Model Token Pricing Matrix (Direct API)

Model Identifier	Input Rate (Cache Miss)	Input Rate (Cache Hit)	Output Rate / 1M	Context Limit	Architectural Specialty
DeepSeek-V3 (Base / Chat)	$0.14 / 1M	$0.014 / 1M (90% off)	$0.28 / 1M	64,000 tokens	General intelligence & coding (671B MoE)
DeepSeek-R1 (Reasoning)	$0.55 / 1M	$0.14 / 1M (75% off)	$2.19 / 1M	64,000 tokens	Mathematical & algorithmic deep reasoning

Official Developer SDK / API Integration Example Python SDK

# DeepSeek AI Direct API Python SDK (OpenAI Compatible)
from openai import OpenAI

client = OpenAI(
    api_key="your_deepseek_api_key",
    base_url="https://api.deepseek.com"
)

# DeepSeek R1 reasoning at $0.55/M in ($0.14 on cache hit) vs $15.00 on o1
response = client.chat.completions.create(
    model="deepseek-reasoner",
    messages=[{"role": "user", "content": "Write an optimal database schema."}]
)
print("Reasoning trace:", response.choices[0].message.reasoning_content)
print("Answer:", response.choices[0].message.content)

Frequently Asked Billing & Pricing Questions

How is DeepSeek R1 able to be 96% cheaper than OpenAI o1?

DeepSeek utilizes Multi-Head Latent Attention (MLA), DeepSeekMoE architecture with 37B active parameters out of 671B, and native FP8 mixed precision training, drastically reducing hardware compute requirements.

Is prompt caching automatic on DeepSeek?

Yes. DeepSeek automatically caches prompt prefixes on the server with zero configuration. When a cache hit occurs, input tokens are billed at $0.14/M instead of $0.55/M on R1.

Can I self-host DeepSeek R1 on my own GPUs?

Yes. DeepSeek R1 weights are published under the MIT license on HuggingFace. You can run quantized versions (1.5B, 7B, 14B, 32B, 70B, or full 671B) on RunPod, Lambda Labs, or local hardware.