DeepSeek Pricing 2026

DeepSeek API list pricing is 80-95% cheaper than GPT-4o and Claude equivalents. At these rates, negotiating further has limited ROI — focus instead on controlling input/output token ratios (shorter prompts, cached context) to cut costs by 30-50% without any vendor negotiation.

DeepSeek supports prompt caching for repeated system prompts and context. Enabling caching on shared system prompts (e.g., same instructions sent with every request) can cut input token costs by up to 90% on high-volume applications.

DeepSeek's direct API automatically applies 50% off for deepseek-chat and 75% off for deepseek-reasoner during off-peak hours (UTC 16:30-00:30). This requires no negotiation — simply schedule non-urgent batch processing, data analysis, or document work during these windows. The deepseek-reasoner output drops from $2.19 to $0.55/1M tokens off-peak.

Cached input tokens cost ~74% less than uncached inputs ($0.07 vs $0.27/1M for deepseek-chat). Keep static content — system prompts, reference documents, few-shot examples — at the beginning of every request. The system automatically caches inputs of 64+ tokens and handles it transparently; no code changes are needed beyond prompt structure.

deepseek-reasoner (R1) costs 2-4x more than deepseek-chat (V3) per token and adds chain-of-thought overhead. For code generation, content writing, data extraction, and general chat, deepseek-chat delivers strong results at a fraction of the price. Reserve R1 only for tasks explicitly requiring multi-step reasoning or math.

Adding explicit output constraints to prompts (e.g., 'Show formula and result only' instead of an open-ended question) reduces output token usage by over 50% in documented examples. For R1, this also shortens chain-of-thought, cutting the most expensive token type. Structure every prompt to request the minimum viable output.

The same DeepSeek model can cost anywhere from free (Google Vertex, CentML during their free tiers) to $5.50/1M on SambaNova. Before integrating into production, benchmark at least three providers on OpenRouter or directly. For experimentation and batch runs, free provider tiers can entirely eliminate inference costs.

Direct API pricing is consistently lower than third-party router markups, especially combined with off-peak discounts. Community consensus is that 'no provider can compete in pricing against DeepSeek's off peak hours' on the direct API. For sustained production usage, direct API access provides the best cost floor.