Foundational 671B Mixture-of-Experts (MoE) model setting the benchmark for low-cost intelligence
deepseek-chat| Scenario | Input Tokens | Output Tokens | Uncached Cost | With Prompt Caching |
|---|---|---|---|---|
| Short Chat Query | 1,000 | 500 | $0.00028 | $0.00015 |
| Document Summarization | 10,000 | 2,000 | $0.00196 | $0.00070 |
| Codebase & Context Analysis | 100,000 | 20,000 | $0.0196 | $0.00700 |
| Batch Corpus Processing | 1,000,000 | 100,000 | $0.1680 | $0.0420 |
Ideal for production workloads demanding flagship capabilities, deep context depth (128k tokens), and reliability from DeepSeek. Excellent when predictable tokenomics and prompt caching support are paramount.
If your use-case requires sub-second streaming latency or ultra-high frequency classification at micro-cent pricing, consider lighter budget options such as Gemini Flash-Lite or Claude Haiku. For deep formal logic, consider dedicated reasoning models like o3.
For DeepSeek-V3, 1 million input tokens costs $0.14, while 1 million output tokens costs $0.28. If using prompt caching, repetitive input prefixes are discounted to $0.014 per million.
DeepSeek-V3 features a maximum context window of 128,000 tokens (~96,000 words), with a maximum output limit of 8,192 tokens per completion.
DeepSeek-V3 utilizes the DeepSeek Byte-level BPE (~128k vocabulary). Token counting on TokenMath runs client-side to ensure maximum privacy.
Yes. DeepSeek-V3 supports prompt caching with a cached input rate of $0.014/1M (saving up to 90% on repeated input context).