DeepSeek
The Chinese lab that wiped out roughly $600 billion of Nvidia's market value in a single day, just by shipping a cheap, open model.
What is DeepSeek?
DeepSeek was founded in Hangzhou in 2023 by Liang Wenfeng, who'd made his money running a quantitative hedge fund rather than coming up through Silicon Valley. That background matters, because it shows in how the company operates: DeepSeek never raised venture capital, trained its models at a fraction of what US labs spent, and gave much of the resulting work away for free under open licenses. Nobody outside China paid much attention until January 2025, when the release of its DeepSeek-R1 reasoning model — reportedly trained for a few million dollars, competitive with OpenAI's best at the time — briefly became the biggest story in tech. It shot to the top of the US App Store within days, and on January 27, 2025, Nvidia's stock dropped 17% in a single session, erasing close to $600 billion in market value as investors questioned whether frontier AI really required the spending levels US labs had assumed.
The product itself is straightforward: a free chat app and website, built on models that are genuinely fast and capable for their price, with no image or video generation and no built-in web search — DeepSeek is a text specialist, not a do-everything assistant. Where it stands out most is the API: DeepSeek's per-token pricing runs roughly 35 to 100 times cheaper than comparable models from OpenAI or Anthropic, and several model versions are released as open weights that anyone can download and run themselves.
The Nvidia figure is widely reported financial-market history from January 27, 2025 (Reuters, Britannica). Model and pricing details reflect DeepSeek's official documentation as of July 2026.
Key features
Extremely low API cost
Tens to a hundred times cheaper per token than equivalent frontier models from US labs.
Open-weight models
Several versions are released under the MIT license, so they can be downloaded and self-hosted at no per-token cost.
Reasoning mode
A "thinking" mode for harder problems, toggled on by default and billed as part of the output.
Strong coding performance
Competitive results on coding benchmarks, and works out of the box with tools like Cursor and Windsurf.
Cache-hit discounts
Repeated prompt prefixes are billed at a steep discount, cutting real-world costs even further.
OpenAI-compatible API
Drop-in compatible endpoint, so switching an existing integration over is mostly a base-URL change.
Pricing
Free chat app
- Full access to current models
- No monthly plans to compare or upgrade
- Text only — no image or video generation
API — V4 Flash
- Budget model for most use cases
- Cache-hit input priced far lower still
- 1M-token context window
API — V4 Pro
- Higher-capability model for harder tasks
- Same 1M-token context window
- Pay-as-you-go, no monthly commitment
There is no paid consumer subscription — DeepSeek monetizes exclusively through API usage. Several models can also be self-hosted for free under an MIT license. Prices reflect DeepSeek's official API pricing as of July 2026 and may change.
Available models
Integrations & platforms
Pros, cons & best for
Pros
- Dramatically cheaper than any comparable API
- Genuinely open-weight models you can self-host
- Strong coding and reasoning performance for the price
Cons
- No image, video, or audio generation at all
- No built-in web search or browsing tools
- Data handling and hosting in China is a dealbreaker for some organizations
Best for
- Developers optimizing for API cost at scale
- Teams wanting to self-host an open-weight model
- High-volume text tasks like classification or extraction
Take a look inside
Alternatives
If you need multimodal features or want a model hosted outside China:
Our verdict
Judged purely on capability, DeepSeek is a legitimately strong reasoning and coding model that happens to cost a fraction of its rivals — the January 2025 market reaction wasn't hype, it was a real reassessment of what frontier AI has to cost. As a daily chat app it's more limited than ChatGPT or Gemini, with no image generation and no web browsing built in, so it's best understood as a specialist rather than a general replacement. For developers optimizing an API bill, or anyone who wants to self-host an open-weight model instead of renting one, it's hard to beat on price. For everyday consumer use, weigh the missing multimodal features and China-based data handling against how much you actually need those things.
FAQ
Is DeepSeek really free?
The chat app and website are free with no paid consumer subscription. DeepSeek makes money through API usage, billed per token, which developers and businesses pay for separately.
Why did DeepSeek cause Nvidia's stock to crash?
In January 2025, DeepSeek released a reasoning model, R1, that performed competitively with leading US models despite reportedly costing far less to train. That challenged assumptions about how much compute frontier AI required, and Nvidia's stock fell 17% in a single day as a result.
Can DeepSeek generate images or video?
No. DeepSeek's models are text-only, with no built-in image, video, or audio generation.
Can I run DeepSeek's models myself?
Yes, several versions are released under an MIT license and can be downloaded and self-hosted, most commonly via Hugging Face, at no per-token cost beyond your own compute.
Is DeepSeek safe to use for sensitive data?
DeepSeek is a Chinese company subject to Chinese data regulations, which is a legitimate consideration for organizations with strict data-residency or compliance requirements — worth weighing before sending sensitive information through it.
Who owns DeepSeek?
DeepSeek is owned by High-Flyer, a Chinese quantitative hedge fund, and was founded by High-Flyer's co-founder, Liang Wenfeng.