DeepSeek

DeepSeek is the lab that made frontier-class reasoning cheap. Its DeepSeek-V4 family delivers quality close to the best Western models at roughly a tenth of the price, and the weights for much of the lineup are public. If your bottleneck is inference cost rather than the last few points of benchmark score, DeepSeek deserves a serious look.

What is DeepSeek?

DeepSeek is a Chinese AI research company known for training highly efficient mixture-of-experts models and releasing open weights alongside a hosted API. As of August 2026 the hosted lineup is three models: deepseek-v4-pro (serving DeepSeek-V4-Pro-0813, generally available), deepseek-v4-flash (serving V4-Flash-0731, in public beta) and deepseek-v4-flash-vision-exp, an experimental image-understanding variant. Both text models carry a 1M-token context window. The consumer chat app is free and unmetered for normal use, which is how most people first encounter the models.

Key Features

• DeepSeek-V4-Pro and V4-Flash: A two-tier lineup covering hard reasoning and cheap throughput, both with 1M-token context.
• Off-peak pricing: API rates roughly halve outside peak hours (peak is 01:00–04:00 and 06:00–10:00 UTC), which is a meaningful lever for batch workloads.
• Aggressive caching: Cache-hit input drops to about $0.007 per million tokens, making repeated-context workloads almost free.
• Open weights: Much of the family is downloadable and self-hostable, with a large ecosystem of community fine-tunes and quantisations.
• Free consumer app: Web and mobile chat with the current models at no cost.

Who Should Use DeepSeek?

DeepSeek is the default for developers running high-volume inference on a budget, for teams that need to self-host on their own hardware, and for anyone whose product economics do not survive frontier API pricing. Researchers and hobbyists benefit from the open weights. The obvious constraint is jurisdiction: the hosted API processes data in China, so regulated industries and many Western enterprises will need to self-host the open weights instead of calling the API.

Pricing

The chat app and web interface are free. On the API, after the 16 August 2026 pricing update, deepseek-v4-flash costs $0.22 per million input tokens off-peak ($0.44 at peak) and $0.66 per million output ($1.32 at peak); deepseek-v4-pro costs $0.66/$1.98 off-peak and $1.32/$3.96 at peak. Cache hits bill at roughly $0.007 per million input tokens. Open weights carry no licence fee — you pay only for the hardware you run them on.

Pros and Cons

Bottom Line

DeepSeek is the price benchmark the rest of the industry now has to answer. For cost-sensitive production workloads, batch processing and self-hosted deployments it is frequently the rational choice, and the free chat app is a genuinely capable everyday assistant. Enterprises with data-residency requirements should plan on running the open weights themselves rather than using the hosted API — which, unusually for a frontier-adjacent lab, is a real option.