← Back to articles Blog

DeepSeek: The Open-Source LLM That Shocked the AI Industry on Cost

In early 2025, a Chinese AI lab called DeepSeek released R1 — a reasoning model that matched OpenAI o1 on key benchmarks while reportedly costing under $6 million to train. For context, GPT-4 was estimated to cost over $100 million.

The AI industry had an immediate reckoning. If frontier performance could be achieved at 1/20th the cost, what did that mean for the entire infrastructure build-out narrative?

What makes DeepSeek different

DeepSeek's efficiency came from several architectural innovations:

• Mixture of Experts (MoE) — only a subset of parameters are active per token, reducing compute without reducing model capacity
• Multi-head Latent Attention (MLA) — reduces KV cache memory requirements dramatically
• GRPO training — a reinforcement learning approach that achieves strong reasoning without massive supervised fine-tuning datasets
• FP8 training — lower precision arithmetic that reduces GPU memory requirements

DeepSeek models available today

Model	Type	Best for	Open source
DeepSeek-V3	Dense MoE	General tasks, coding	✓ MIT
DeepSeek-R1	Reasoning	Math, logic, code	✓ MIT
DeepSeek-R1-Distill	Small reasoning	Local deployment	✓ MIT

DeepSeek pricing vs frontier models

DeepSeek's API is significantly cheaper than OpenAI or Anthropic:

Model	Input per 1M tokens	Output per 1M tokens
DeepSeek-V3	$0.27	$1.10
DeepSeek-R1	$0.55	$2.19
GPT-4o	$5.00	$15.00
Claude Sonnet	$3.00	$15.00

At these prices, DeepSeek is 10-18x cheaper than frontier alternatives on input tokens. The catch: data privacy concerns, since the API sends data to Chinese servers. For many enterprise use cases, self-hosting the open-source weights is the preferred approach.

How to use the DeepSeek API

The DeepSeek API is OpenAI-compatible. Sign up at platform.deepseek.com to get an API key, then use it with any OpenAI SDK:

from tokoscope import wrap
from openai import OpenAI

client = wrap(OpenAI(
    base_url="https://api.deepseek.com/v1",
    api_key="YOUR_DEEPSEEK_KEY"
), api_key="ts_live_...")

# Standard (non-thinking) call
response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Explain merge sort"}]
)

# Thinking mode — deeper reasoning, more output tokens
response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Solve this math problem step by step..."}],
    extra_body={"thinking": {"type": "enabled", "budget_tokens": 8000}}
)

Current model strings as of July 2026:

• deepseek-v4-flash — main model, non-thinking mode (formerly deepseek-chat)
• deepseek-v4-pro — highest quality, 3x cost
• Note: deepseek-chat and deepseek-reasoner were retired July 24, 2026. Update any existing integrations.

The real lesson: efficiency over scale

DeepSeek proved that intelligent architecture beats raw compute spend. The same principle applies to your production LLM costs — the teams spending the least per output quality are the ones who optimize prompt efficiency, implement semantic caching, and track token usage at the endpoint level.