DeepSeek

Overview

DeepSeek is a Chinese AI research lab that made global headlines with its open-source language models — DeepSeek-V3 (a 671B-parameter Mixture-of-Experts model with 37B active per token) and DeepSeek-R1 (a reasoning model that matched OpenAI-o1 on math, code and logic at a fraction of the cost). As of 2026 the API runs on DeepSeek V4, offering an OpenAI-compatible endpoint plus an Anthropic-format endpoint, with some of the lowest per-token prices in the industry. Models and code are released under the MIT License with commercial use and distillation permitted.

Key Features

• DeepSeek V4 generation (current): two API models — deepseek-v4-flash (1M-token context, up to 384K max output, 2,500 concurrent, the default for web/chat) and deepseek-v4-pro (higher quality, 500 concurrent). Both support thinking and non-thinking modes, JSON output, tool calls, Chat Prefix and FIM completion (beta).
• DeepThink mode: a toggle on chat.deepseek.com (and the API) switches V4 between fast non-thinking answers and step-by-step reasoning — the successor to R1-style chain-of-thought.
• OpenAI + Anthropic compatible: use the standard OpenAI SDK (base URL ) or Anthropic-format clients with a one-line model swap; works as a backend for Claude Code, GitHub Copilot, OpenCode and similar agent tools.
• Industry-low pricing: ~90% cheaper than flagship OpenAI/Anthropic models; prompt caching (90% discount on cache hits) and JSON mode included.
• Open-source ecosystem: V3 and R1 weights are MIT-licensed (code) with commercial use and distillation allowed; distilled R1-Distill models (Qwen 1.5B–32B, Llama 8B/70B) run on modest hardware. V3 trained on 14.8T tokens with an MoE + MLA architecture for efficient inference.
• Responses API: currently supported on deepseek-v4-flash, with deepseek-v4-pro support expected early August 2026.

Pricing

Per 1M tokens (2026 V4 rates; DeepSeek has announced a significant API price increase is planned — verify live rates before capacity planning):

• deepseek-v4-flash: input $0.14 (cache miss) / $0.0028 (cache hit), output $0.28.
• deepseek-v4-pro: input $0.435 (cache miss) / $0.003625 (cache hit), output $0.87.
• Legacy rates for reference: DeepSeek-V3.1 $0.27 in / $1.10 out; DeepSeek-R1 $0.55 in / $2.19 out (reasoning tokens billed at output rates). The legacy deepseek-reasoner ID routes to v4-flash thinking mode and is fully retired on July 24, 2026 at 15:59 UTC.
• Context caching reduces input cost by ~90% on cache hits; off-peak discounts were discontinued in September 2025.

Best For

Developers and businesses building AI products, chatbots, coding assistants and content pipelines that need frontier-quality or near-frontier reasoning at dramatically lower cost — and teams that want full control via open weights for self-hosting, fine-tuning and distillation.

FAQ

Is DeepSeek open source?

Yes — the V3 and R1 models and code are released under the MIT License (model weights under a separate but permissive Model License), allowing commercial use, modifications and distillation.

How much does the DeepSeek API cost?

As of 2026, deepseek-v4-flash is $0.14/M input (or $0.0028/M on cache hit) and $0.28/M output; deepseek-v4-pro is $0.435/M input and $0.87/M output — about 90% cheaper than comparable OpenAI/Anthropic models. A price increase has been announced.

Is it OpenAI-compatible?

Yes — the API mirrors OpenAI's Chat Completions schema (and also offers an Anthropic-format endpoint), so existing SDKs work by changing only the base URL and model name.

What happened to R1?

DeepSeek-R1's reasoning is now delivered via V4's thinking mode. The legacy deepseek-reasoner API ID routes to deepseek-v4-flash in thinking mode and retired on July 24, 2026; the open R1 weights remain available for self-hosting.