Build with
DeepSeek API.

Sign up at platform.deepseek.com with email or Google — no credit card required. Every new account receives 5 million free tokens automatically. Install the OpenAI SDK (pip install openai), set base_url="https://api.deepseek.com/v1" and your DEEPSEEK_API_KEY. Use model deepseek-v4-flash for general use. The full API documentation is at api-docs.deepseek.com.

The legacy aliases deepseek-chat and deepseek-reasoner are deprecated and will stop working at July 24, 2026 15:59 UTC — with no grace period and no silent fallback. Calls will return errors. The migration is a one-line change: update model to deepseek-v4-flash (replacing deepseek-chat) or deepseek-v4-pro (replacing deepseek-reasoner). The base URL, authentication, and request format are identical. Search your entire codebase including retry logic, fallback handlers, and configuration files — not just primary API calls.

DeepSeek's caching is fully automatic — no code changes, no explicit cache keys, no SDK flags. Any prompt prefix that has been seen recently will be served from cache at a 90% discount ($0.014/1M vs $0.14/1M for V4-Flash). To maximize cache hits: (1) Put your system prompt and any fixed context at the very top of every request. (2) Keep the variable parts (user query, latest message) at the end. (3) Use usage.prompt_cache_hit_tokens in the response to monitor your actual cache hit rate. A well-designed production app with a 1,000-token system prompt on 100K daily requests can save hundreds of dollars monthly purely from caching.

V4-Flash: 284B MoE (13B active), $0.14/$0.28 per 1M, 83 tok/s, 79% SWE-bench. The recommended default for most production workloads — 12.4× cheaper per output token than Pro. Start here. V4-Pro: 1.6T MoE (49B active), $1.74/$3.48 per 1M (regular, $0.435/$0.87 promo until May 31), ~40 tok/s, 80.6% SWE-bench, Codeforces #1, IMO Gold with Think Max. Use when quality matters more than cost — complex multi-step agents, hard reasoning, or benchmark-critical tasks. Best practice: benchmark both on your actual task before committing. For most use cases, Flash is sufficient.

Pass the extra_body parameter to control reasoning effort per request. Three modes: (1) {"thinking": {"type": "disabled"}} — Non-Think, instant response, no CoT. (2) {"thinking": {"type": "enabled", "budget": "high"}} — Think High, structured analytical reasoning. (3) {"thinking": {"type": "enabled", "budget": "max"}} — Think Max, full reasoning chain, most tokens, best quality for hard problems. When using thinking modes, the response contains both content (the final answer) and reasoning_content (the chain-of-thought). Important: In multi-turn conversations, only include content in the history — never reasoning_content. Including the reasoning chain in history degrades future responses.

DeepSeek does not publish fixed public rate limits — limits are applied dynamically based on account status and server load. The free tier (5M tokens) is subject to stricter rate limiting than paid accounts. During peak demand (especially during model launches), the direct API may experience capacity constraints. For mission-critical production systems requiring 99.9%+ availability SLAs, DeepSeek recommends routing through cloud providers like AWS Bedrock or Azure AI, or inference providers like Together AI or Fireworks, which offer dedicated capacity at a modest premium (~25% higher per-token cost).

Historically, DeepSeek offered 50–75% additional discounts for API calls made during 16:30–00:30 UTC on legacy V3 and R1 models. As of the V4 launch (April 24, 2026), the current official documentation shows the standard pricing table without an explicit off-peak tier listed. The 75% promotional discount on V4-Pro (until May 31, 2026) effectively provides a similar cost reduction for Pro users. Always check the official pricing page at api-docs.deepseek.com/quick_start/pricing for the most current rates — DeepSeek adjusts pricing and promotions regularly.

Yes. DeepSeek models are available through multiple cloud providers for enterprises requiring data residency, compliance guarantees, or higher reliability SLAs: AWS Bedrock (HIPAA/SOC2 eligible, data residency in your region), Azure AI (GDPR-compliant EU region available), Together AI and Fireworks (faster US inference, OpenAI-compatible API). Cloud provider pricing is typically 20–30% higher per token than the direct DeepSeek API, but includes enterprise SLAs, compliance certifications, and dedicated capacity. For sensitive data (healthcare, financial, legal), cloud provider routing is strongly recommended over the direct DeepSeek API.