DeepSeek V4 is here. MIT open-source license, a 1 million-token context window, and output at just $3.48 per million tokens — roughly one-tenth of GPT-5.5. Better yet, hands-on tests show its coding ability beating Claude Opus 4.6 and GPT-5.4. If you've been thinking about migrating away from Claude or GPT, this is the best window to do it.

What Is DeepSeek V4

The DeepSeek V4 series includes two models:

Note: The above are standard API prices. On cache hits, V4-Pro input drops to $0.145 and V4-Flash input drops to $0.02. V4-Pro is currently 75% off for a limited time (through May 5), and the cache-hit price is a permanent price cut.

Price comparison

The cost of processing 1 million input tokens + 1 million output tokens:

• GPT-5.5: $35 (standard API)
• Claude Opus 4.7: $30
• DeepSeek-V4-Pro: $5.22 (cache miss) / $3.625 (cache hit)
• DeepSeek-V4-Flash: $0.42 (cache miss) / $0.308 (cache hit)

API pricing comparison of mainstream models — DeepSeek V4-Flash costs less than 2% of GPT-5.5

Before You Start

• A DeepSeek API account (sign up at platform.deepseek.com)
• Your coding tool (Claude Code, Cursor, Aider, etc.)
• Estimated migration time: 10-30 minutes

Step by Step

Step 1: Get your API key

1. Log in to the DeepSeek open platform
2. Go to the API Keys page and create a new key
3. Copy and save it (it is shown only once)

Step 2: Configure the DeepSeek endpoint in your coding tool

Tech blogger Sean Donahoe shared his migration experience: no OpenRouter, no proxy — just use the DeepSeek native API directly.

Claude Code configuration example:

# Set the DeepSeek API endpoint
export DEEPSEEK_API_KEY="your-api-key"
export DEEPSEEK_BASE_URL="https://api.deepseek.com"

# Configure a custom model endpoint in Claude Code
claude config set model deepseek-v4-pro
claude config set apiBase https://api.deepseek.com
claude config set apiKey $DEEPSEEK_API_KEY

Cursor configuration:

Open Settings -> Models -> OpenAI API Key, enter your DeepSeek API Key and Base URL https://api.deepseek.com, and select the model deepseek-v4-pro.

Tip: If you use several tools at once, a good approach is to run V4-Flash for everyday lightweight tasks (it's cheaper) and save V4-Pro for complex coding and reasoning work.

Step 3: Verify your cache hit rate

DeepSeek V4's cache hit rate is very high (about 95% for V4-Pro and 91% for V4-Flash), which means most tokens are billed at the cache-hit price.

# Check your API usage and cache hit rate
import requests

response = requests.get(
    "https://api.deepseek.com/usage",
    headers={"Authorization": f"Bearer {API_KEY}"}
)
print(response.json())

Step 4: Compare results

We recommend running the same task on both models first to compare results and costs.

QbitAI's real-world numbers: a coding job of 35 million tokens cost RMB 31.73 before the discount and just RMB 5.34 after — an overall saving of about 83%.

DeepSeek V4 cost comparison before and after the discount — an overall saving of 83%

Technical highlights

• 1M-token context window: Up from 128k in V3 to 1M, well suited to working with large codebases
• MIT open-source license: Self-deploy to bypass API billing, ideal for compliance-sensitive scenarios
• 90% smaller KV cache: Inference efficiency improves dramatically in million-token context scenarios
• Support for Huawei Ascend NPUs: Inference traffic may eventually run on domestic Chinese compute

FAQ

• How big is the gap between V4 and GPT-5.5?: DeepSeek officially acknowledges it is still about 3-6 months behind the frontier models, but in real-world everyday coding tasks it performs on par or better
• When does the Pro limited-time discount end?: Through May 5 at 23:59; the cache-hit input price cut is permanent
• What setup does self-deployment need?: V4-Flash's smaller parameter count can be deployed on a single multi-GPU machine; V4-Pro's 1.6T parameters require a cluster