DeepSeek

DeepSeek is an AI lab spun out of High-Flyer, a Chinese quant trading firm. Their flagship products are DeepSeek V3 (a general-purpose chat model) and DeepSeek R1 (a reasoning-focused model), both available through the free chat interface at deepseek.com, the paid API, and as open-source weights on Hugging Face.

The positioning is straightforward: frontier-grade capability at a fraction of the cost. DeepSeek doesn't try to compete with ChatGPT on polish or Claude on writing quality — they compete on price-performance, and on that axis they win decisively.

V3 is DeepSeek's flagship general-purpose model. It handles chat, coding, knowledge questions, and analysis competently. Quality is in the GPT-4 class — meaningfully behind GPT-5 and Claude Sonnet 4.6 on the hardest tasks, but plenty good enough for everyday use.

The practical experience using V3 in the chat product is fine but unremarkable. Responses are fast. The interface is functional. There's no voice mode, no image generation, no fancy artifacts. It's a chat box that gets you a competent answer. For free, you can't really complain.

R1 is the model that made DeepSeek famous. Unlike V3, R1 is a reasoning model — it thinks step-by-step through problems before answering, similar to OpenAI's o1 or DeepSeek's own competition for that category.

On math benchmarks, logic puzzles, and multi-step reasoning tasks, R1 is genuinely competitive with the best Western reasoning models. On a price-per-token basis, R1 is roughly 1/20th the cost of OpenAI's o1 equivalent. For developers building reasoning-heavy applications, the math is overwhelming.

DeepSeek's API pricing is the real story. V3 is $0.14 per million input tokens and $0.28 per million output tokens. With cached input, the input price drops to $0.014 per million — orders of magnitude cheaper than GPT-5 or Claude Sonnet 4.6.

For developers operating at scale, this changes the unit economics of AI features. A product feature that costs $1000/month on GPT-5 might cost $50/month on DeepSeek V3. Whether the quality difference justifies the price difference depends on the use case, but the fact that the choice exists is significant.

DeepSeek releases their model weights openly under a permissive license. You can download V3 and R1 from Hugging Face, run them on your own GPUs, and have full control over the deployment. The licensing terms allow commercial use.

In practice, very few organizations actually self-host — V3 is over 600B parameters and requires meaningful infrastructure. Most companies who want self-hosting end up using a hosted alternative like Together AI, Fireworks, or Replicate that runs the open weights. Even those hosted alternatives are dramatically cheaper than proprietary frontier models.

For Western businesses, the Chinese hosting of DeepSeek's hosted product is a legitimate concern. If you process EU customer data (GDPR), US healthcare data (HIPAA), regulated financial data, or competitive IP, sending it to Chinese infrastructure introduces compliance and security risks that most legal teams will not approve.

For non-sensitive use cases — personal coding, brainstorming, learning — the hosted product is fine. For anything sensitive, either self-host the open weights, use a Western re-host like Together AI, or pick a Western alternative.

This is the single biggest reason DeepSeek hasn't fully replaced GPT or Claude in enterprise: the legal blocker is real.

DeepSeek's hosted product operates under Chinese regulations that require AI services to align with state guidelines. Questions about politically sensitive topics — Tiananmen Square, Taiwan, Hong Kong protests, Xinjiang, and similar — produce refused or evasive responses. The self-hosted open weights have less of this baked in but still show some alignment artifacts.

For most professional use cases this doesn't matter. For journalists, researchers, or anyone working on China-related topics, this is a real limitation.

DeepSeek vs. ChatGPT/Claude: Western models win on polish, ecosystem, and the censorship-free baseline. DeepSeek wins on cost — by a lot. Pick based on whether your use case is cost-sensitive or quality-sensitive.

DeepSeek vs. Qwen: Both are Chinese open-source frontier models. Similar capability, similar pricing, similar privacy considerations. Qwen has slightly stronger Chinese-language performance; DeepSeek has the higher-profile reasoning model.

DeepSeek vs. Llama (Meta): Llama is the strongest Western open-source alternative. Similar parameter scale, similar performance band, dramatically different geopolitical context. For organizations that need open-source AI but can't use Chinese-hosted infrastructure, Llama is the obvious choice.

DeepSeek is genuinely impressive AI at a price that breaks the rest of the market. For developers building on a budget, the API is the right starting point — V3 for general chat, R1 for reasoning-heavy tasks. The chat product is solid but unremarkable; the real value is what enterprises and developers can do with the API economics. The Chinese hosting question is the gating factor for serious commercial use; resolve that before committing.