DeepSeek V4: The Open-Weights Model
That Changed AI Coding Forever

Released April 24, 2026. V4-Pro scores 80.6% SWE-bench at $0.87/M output — that's 28x cheaper than Claude Opus 4.8. Here's everything you need to know.

80.6%
SWE-bench Verified
$0.87
Per 1M Output Tokens
1M
Context Window
MIT
Open Source License

V4-Pro vs V4-Flash: Which One Should You Use?

DeepSeek shipped two models on the same day. V4-Pro is the frontier model; V4-Flash is the budget speed demon. Both support 1M-token context and 384K max output.

🚀 V4-Pro

$0.87/1M output tokens
Input: $0.435/M · Cache hit: $0.004/M
  • 1.6 trillion total parameters
  • 49B active parameters per token
  • 80.6% SWE-bench Verified
  • Best for: complex coding, multi-file refactoring
  • 28.7x cheaper than Claude Opus 4.8
  • 34.5x cheaper than GPT-5.5

⚡ V4-Flash

$0.28/1M output tokens
Input: $0.14/M · Cache hit: $0.003/M
  • 284 billion total parameters
  • 13B active parameters per token
  • 79.0% SWE-bench Verified
  • Best for: fast iteration, daily coding tasks
  • ~89x cheaper than Claude Opus 4.8
  • Replaces legacy deepseek-chat endpoint

Pricing Comparison: DeepSeek V4 vs The Competition

How much does a dollar buy you in output tokens? This is where DeepSeek V4 dominates.

ModelOutput Price (per 1M)Tokens per $1SWE-bench
DeepSeek V4-Flash$0.283,571,42879.0%
DeepSeek V4-Pro$0.871,149,42580.6%
DeepSeek Chat (legacy)$1.10909,09067.8%
Gemini 3.1 Pro$10.5095,23880%+
GPT-5.5$30.0033,333~87%
Claude Opus 4.8$25.0040,00088.6%

The verdict: A $1 budget gets you 1.15M tokens from V4-Pro vs just 40K from Claude Opus. For most coding tasks, that's an easy tradeoff.

How to Set Up DeepSeek V4

With Claude Code (Terminal)

DeepSeek V4 supports the Anthropic API format natively — just change the base URL:

# 1. Get your API key from platform.deepseek.com
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_API_KEY=sk-your-deepseek-key

# 2. Use claude-code normally — it routes through DeepSeek V4
claude-code "refactor the authentication module"

# To switch between Pro and Flash:
export DEEPSEEK_MODEL=deepseek-v4-pro
# or
export DEEPSEEK_MODEL=deepseek-v4-flash

With Cursor IDE

# Settings → Models → Add Custom Model
# API Base URL: https://api.deepseek.com/v1
# Model: deepseek-v4-pro (or deepseek-v4-flash)
# API Key: your DeepSeek API key

With OpenRouter (Recommended for Multi-Model)

# OpenRouter gives you DeepSeek V4 + Claude + GPT in one key
export OPENROUTER_API_KEY=sk-or-your-key
# Model: deepseek/deepseek-v4-pro
# Model: deepseek/deepseek-v4-flash

# Works with Claude Code, Aider, Continue.dev, Cline

Self-Hosting (MIT License)

V4 weights are on Hugging Face under MIT license. V4-Flash (284B/13B active) can run on a single 8xH100 node. V4-Pro (1.6T/49B active) needs multi-node inference. Use vLLM or SGLang for serving.

Benchmark Deep Dive

BenchmarkV4-ProV4-FlashClaude Opus 4.8GPT-5.5
SWE-bench Verified80.6%79.0%88.6%~87%
Chatbot Arena Elo~14501433~1500+~1490
Coding Arena Elo1479
Context Window1M1M200K (1M enterprise)128K
Max Output384K384K64K16K
LicenseMIT (open)MIT (open)ProprietaryProprietary

💡 Our Take

DeepSeek V4 is the best value proposition in AI coding today. For 90% of coding tasks, V4-Pro delivers results indistinguishable from models costing 28x more. Use Opus or GPT-5.5 for the hardest 10% of tasks. For everything else, V4 is the smart play.

Try DeepSeek V4 Today

Start with $5 free credit on OpenRouter — access DeepSeek V4, Claude, GPT-5, and 200+ models with one API key.

Get Started on OpenRouter →

Buy Crypto to Pay for API Credits

Many developers pay for API credits with crypto. Get the lowest fees on Binance.

Sign Up on Binance (Fee Discount) →

Frequently Asked Questions

What is DeepSeek V4?
DeepSeek V4 is an open-weights AI model family released April 24, 2026 by DeepSeek (a Chinese AI lab). It includes V4-Pro (1.6T total params, 49B active, $0.87/M output) and V4-Flash (284B total, 13B active, $0.28/M output). Both support 1M-token context windows and 384K max output tokens. V4-Pro scores 80.6% on SWE-bench Verified, matching frontier models at 28x lower cost.
How much does DeepSeek V4 cost compared to Claude and GPT?
V4-Pro costs $0.87/M output tokens — 28.7x cheaper than Claude Opus 4.8 ($25/M) and 34.5x cheaper than GPT-5.5 ($30/M). V4-Flash is even cheaper at $0.28/M output. Cache hits (for repeated prompts) are nearly free: $0.004/M for V4-Pro. In practical terms, $1 buys 1.15M output tokens from V4-Pro vs just 40K from Claude Opus.
Can I use DeepSeek V4 with Claude Code or Cursor?
Yes! DeepSeek V4 supports the Anthropic API format natively. For Claude Code, just set ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic and your DeepSeek API key. For Cursor, add DeepSeek as a custom model provider in Settings → Models. You can also access it through OpenRouter, which lets you switch between DeepSeek, Claude, and GPT with one key.
Is DeepSeek V4 really open source?
Yes. The weights are released under the MIT License on Hugging Face. You can self-host, fine-tune, and use commercially without restrictions. The models use Mixture-of-Experts (MoE), so while total parameters are massive, active parameters per token are manageable. V4-Flash can run on a single 8xH100 node.
When will DeepSeek V4 replace the old deepseek-chat model?
The legacy deepseek-chat and deepseek-reasoner endpoints now map to deepseek-v4-flash. They will be fully retired after July 24, 2026. If you're using the old endpoint, your requests already route to V4-Flash automatically. But you should update your code to use the new endpoint names for best results.
What's the catch? Why is DeepSeek so cheap?
DeepSeek is funded by High-Flyer Capital, a major Chinese quantitative hedge fund. Their MoE architecture is extremely efficient — only 49B (Pro) or 13B (Flash) parameters activate per token, even though total parameters are much larger. They also benefit from lower labor costs in China and aggressive pricing to capture market share. The MIT license means you can self-host if you prefer not to depend on their API.

Related Guides