DeepSeek

What is DeepSeek?

DeepSeek is an AI lab from Hangzhou, funded by the hedge fund High-Flyer. The big bang came in January 2025 when its R1 model matched o1-level reasoning at a fraction of US training costs and knocked double digits off Nvidia's stock in a day. Since then DeepSeek has shipped relentlessly. As of August 2026 the current generation is V4 (April 2026): V4 Pro with 1.6 trillion parameters (49B active) and the smaller V4 Flash with 284 billion (13B active). Both carry a 1 million token context window and ship as open weights under MIT license.

Is DeepSeek free?

Yes, entirely. The chat at chat.deepseek.com and the mobile app have no paid tier, no Plus, no Pro, no paywall. That sets DeepSeek apart from practically every Western provider. Developers additionally get a 5 million token starter grant for the API.

API pricing: the real lever

V4 Flash costs $0.14 per million input tokens and $0.28 output, V4 Pro $0.435 and $0.87. Cache hits push input cost down drastically. For comparison: frontier models from OpenAI and Anthropic run ten to fifty times higher. For high-volume workloads like classification, extraction or chat backends this is the cheapest route to serious quality.

The catch: data protection

DeepSeek processes chat data on servers in China. Fine for casual research, a GDPR problem for EU companies: no customer data, no internals in the free chat. To use the models without sending data to China, self-host the open weights or pick an EU provider serving DeepSeek models. That is exactly where the MIT license earns its keep.

How it compares

Against ChatGPT and Claude, DeepSeek wins on price and openness rather than peak quality in every task. For European data sovereignty, Mistral is the closer alternative. What open weights mean in practice is covered in our LLM entry.

Wondering which model belongs in your stack? Our AI consulting answers exactly that, GDPR assessment included.