236B total parameters with only 21B activated per token. 338 programming languages, 128k context, and automatic prompt caching down to $0.014/1M tokens.
Input / 1M (Cache Hit)
$0.014
90% auto-caching discount
Input / 1M (Cache Miss)
$0.14
Uncached prompt rate
Output / 1M Tokens
$0.28
Generated code tokens
Context Window
128,000
Full repo analysis
âš¡ Prompt Caching Impact Simulator
When used in agentic loops (Aider, Cline, Continue.dev), repository contexts and file trees get repeatedly sent. Test how DeepSeek's context cache hit rate cuts your bill:
Total Monthly Input Tokens (Repo context + History)50M tokens
Dense models like Llama 3 70B activate all 70 billion parameters for every single token, requiring massive memory bandwidth.
DeepSeek Coder V2 routes each token to only 21B parameters out of 236B total parameters.
This allows the model to retain specialized knowledge across 338 obscure programming languages and libraries without paying the inference compute cost of a 236B dense model.
Frequently Asked Questions
What is the token pricing for DeepSeek Coder V2?
DeepSeek Coder V2 is priced at $0.14 per 1M input tokens on cache miss, an ultra-low $0.014 per 1M tokens on cache hit (90% discount), and $0.28 per 1M output tokens.
How does DeepSeek Coder V2 achieve such low prices?
DeepSeek Coder V2 uses a 236B Mixture-of-Experts (MoE) architecture where only 21 billion parameters are actively computed per token, dramatically reducing inference compute while maintaining the broad knowledge of a massive model.
How does DeepSeek Coder V2 compare with GPT-4o for coding?
On standard coding benchmarks (HumanEval, MBPP, SWE-bench Lite), DeepSeek Coder V2 scores within 2-4% of GPT-4o while being approximately 94% cheaper on uncached tokens and over 99% cheaper on cached prompts.