DeepSeek V4 API

What DeepSeek V4 does

Use one token for DeepSeek V4 Flash and V4 Pro across high-throughput chat, complex reasoning, tool use, and agentic coding, with OpenAI Chat Completions and Anthropic Messages compatibility.

• Two purpose-built tiers — V4 Flash prioritizes speed, throughput, and cost; V4 Pro targets difficult reasoning, code review, and multi-step agent work. Both support thinking and non-thinking modes.
• 1M-token context window — Built for long codebases, documents, multi-turn tool use, and large-context tasks, with output up to 384K tokens while latency and cost remain workload-dependent.

DeepSeek V4 API features

• Controllable thinking mode — Toggle thinking with thinking.type, select high or max effort, and read reasoning from reasoning_content.
• Tool calls and JSON output — Supports OpenAI-style tools, tool_choice, and response_format for structured workflows and agent loops.
• Complete streaming usage — Set stream_options.include_usage=true to receive total token usage before [DONE].
• Cache-aware usage — Responses separate prompt_cache_hit_tokens and prompt_cache_miss_tokens so long-context cache savings are measurable.

What people build with DeepSeek V4

• Code review and repository agents — Analyze large repositories with 1M context for cross-file changes, test diagnosis, and multi-step tool use.
• Long-document analysis — Process contracts, research, technical specifications, and knowledge bases into structured findings and traceable summaries.
• High-throughput production chat — Use V4 Flash for support, content processing, classification, and batch automation.

Getting started with the DeepSeek V4 API

The endpoint is OpenAI-compatible, so existing client code usually needs only a base URL and a model name.

• Step 1: Create an API token — Create an API token with access to DeepSeek V4.
• Step 2: Choose Flash or Pro — Start with Flash for everyday or high-concurrency work, and use Pro for difficult reasoning, coding, and agents.
• Step 3: Configure thinking — Thinking is enabled by default; explicitly disable it when sampling controls are needed.
• Step 4: Add streaming and tools — Use SSE in production and preserve reasoning_content across subsequent turns that follow tool calls.

DeepSeek V4 API pricing

Token-based billing · see the pricing page for current rates

Why choose CCAPI for DeepSeek V4

• Cache-aware billing — Cache-hit and cache-miss input tokens are identified separately and billed at their corresponding rates.
• Unified auth and usage logs — Manage tokens, access permissions, request logs, and usage accounting without maintaining extra credentials.
• OpenAI-compatible migration — Existing OpenAI SDK clients only need a new base URL, token, and model value.

Other models on the same endpoint

The same API key and base URL reach every model below. Moving from DeepSeek V4 to any of them means changing the model name in the request body.

• Claude Haiku 4.5 API — chat model on the same endpoint
• MiniMax M2.7 API — chat model on the same endpoint
• GLM-5 API — chat model on the same endpoint
• Qwen 3.5 Plus API — chat model on the same endpoint
• Kimi K2.5 (Moonshot) API — chat model on the same endpoint
• GPT-4o API — chat model on the same endpoint

Frequently asked questions