DeepSeek
Listed at https://platform.deepseek.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | DeepSeek-V3: ¥0.14/M token (≈$0.02), R1: ¥0.14/M | per million tokens |
| Output | DeepSeek-V3: ¥0.28/M token (≈$0.04), R1: ¥0.28/M | per million tokens |
🤖 Supported Models (3)
✨ Pros
- ✓Industry-lowest pricing: DeepSeek-V3 ¥0.14/M input
- ✓R1 reasoning model: o1-comparable capabilities at ¥0.14/M
- ✓Direct China access, no proxy needed
- ✓¥500 free credits on signup, generous trial
- ✓Fully open-source model series, active community
- ✓Chinese-optimized, excellent for local use cases
⚠️ Cons
- ×Limited model selection (only 3)
- ×API occasional instability under high China traffic
- ×English capabilities slightly behind GPT-4o/Claude
- ×No multimodal output
- ×Limited Function Calling (R1 doesn't support, V3 limited)
🎯 Best For
China-based developers; cost-sensitive applications; reasoning tasks (R1); bilingual scenarios
💰 Pricing & Plans
| Model | Input (¥/M tokens) | Output (¥/M tokens) | Best For |
|---|---|---|---|
| DeepSeek-V3 | ¥0.14 (≈$0.02) | ¥0.28 (≈$0.04) | General chat, bilingual tasks |
| DeepSeek-R1 | ¥0.14 (≈$0.02) | ¥0.28 (≈$0.04) | Chain-of-thought reasoning |
| DeepSeek-Coder-V2 | ¥0.14 (≈$0.02) | ¥0.28 (≈$0.04) | Code generation, code review |
| Cache hit (V3/R1) | ¥0.014 (≈$0.002) | — | Repeated system prompts, 90% off |
| Off-peak discount | 0.5× input/output | 0.5× input/output | 00:30–08:30 Beijing time, daily |
🔧 API & Developer Experience
- •API Style: OpenAI-compatible /v1/chat/completions endpoint — drop-in replacement for any OpenAI SDK by changing the base URL.
- •Base URL: https://api.deepseek.com/v1 — direct from China, no proxy required.
- •SDK Compatibility: Works with the official OpenAI Python/Node SDKs by swapping base_url and api_key. Community SDKs in Go, Java, Rust.
- •Function Calling / Tools: Supported on DeepSeek-V3 (stable) and on R1 via the dedicated deepseek-reasoner endpoint with tool_choice; JSON-schema validated output.
- •Streaming: Server-Sent Events (SSE) token streaming included by default; R1 also streams the reasoning chain (CoT) before the final answer.
- •Context Window: 128K tokens for V3 and R1 — enough for medium-length documents, long codebases, or extended multi-turn dialog.
- •Prompt Caching: First-party cache hits priced at ¥0.014/M (1/10 of normal input) — repeats with the same system prefix are dramatically cheaper.
🧠 R1 Reasoning at ¥0.14/M
DeepSeek-R1's defining trait is o1-comparable chain-of-thought reasoning at a fraction of the price. Where OpenAI's o1/o3 charge $10–$60 per million tokens for advanced reasoning, R1 ships at ¥0.14 input / ¥0.28 output — roughly 1/40 of the equivalent OpenAI tier. The model exposes its full reasoning trace via streaming, so developers can inspect CoT steps for debugging or audit before the final answer. R1 performs strongly on math (AIME, MATH-500), code reasoning (LiveCodeBench), and graduate-level QA benchmarks where step-by-step logic matters more than fluency. V3 is the cheaper generalist, R1 the reasoning specialist — both priced identically, so the choice is workload-driven. For Chinese-language reasoning tasks, R1 is also a top performer, often matching or beating Western reasoning models on Chinese math olympiad problems.
🌐 Regional Availability & Latency
DeepSeek is one of the few frontier-tier model providers with first-party direct access from mainland China. The api.deepseek.com endpoint serves China-based traffic over domestic BGP routes — typical latency from Beijing/Shanghai is 80–250ms for first-token streaming and 5–15 tokens/sec generation speed, with no VPN or proxy required. International access (api.deepseek.com from US/EU) is also supported via the same endpoint with slightly higher latency. For production workloads, DeepSeek is a natural default for China-serving applications where data residency (no cross-border transfer) and cost predictability matter. The trade-off vs OpenAI/Anthropic is English-language fluency, multimodal capabilities, and ecosystem tooling — DeepSeek-V3 is text-only and the tool-calling surface is narrower than GPT-4o's. Developers building China-first AI products typically pair DeepSeek with international models rather than choosing one over the other.