DeepSeek 

DeepSeek arrived in January 2025 and knocked roughly a fifth off Nvidia’s share price within a week, because it showed that competitive models could be trained for far less than anyone had assumed. Eighteen months later, the picture is calmer, but the core claim holds. The models are genuinely good, the app costs nothing, and the API is priced so far below Western rivals that the comparison feels unfair. What has changed is the surrounding context. Governments in several countries restrict it, the data question is real, and DeepSeek has now warned that its own prices are going up.

What Is DeepSeek?

DeepSeek is both a chat assistant and a family of AI models. The company was founded in Hangzhou in 2023 and grew out of High-Flyer, a quantitative hedge fund, with Liang Wenfeng as founder.

Its first models were coding-focused, released in late 2023. The turning point came in January 2025, when the DeepSeek app launched on iOS and Android, built on the R1 reasoning model. Within a fortnight, it was the most downloaded free app on the US App Store, and the resulting shock to assumptions about training costs moved markets.

What separates DeepSeek from OpenAI, Google, and Anthropic is openness. Model weights have been published under permissive terms, including MIT licensing for R1 and V3, so anyone can download and run them on their own hardware. That is a genuine structural difference, not a marketing line, and it is why DeepSeek matters to teams that cannot send data to a third party.

One point of confusion worth settling: DeepSeek is privately held. There is no DeepSeek stock and no share price to look up.

How Does DeepSeek Work?

In the app, you sign in and type. There is a toggle for thinking mode, which makes the model reason step by step before answering, and a search toggle for current information. Nothing else needs configuring.

The current generation is DeepSeek-V4, released in April 2026 and split into two models. V4-Flash is the fast, cheap workhorse, updated again at the end of July 2026 with a large jump in agent and coding benchmarks. V4-Pro is the heavier model for harder problems. Both handle a one million token context window and both support thinking and non-thinking modes in a single model rather than requiring you to switch products.

For developers, the API accepts both OpenAI and Anthropic request formats, so migrating an existing project is often a base URL change rather than a rewrite. Context caching cuts repeat costs dramatically, with cached input priced roughly fifty times lower than fresh input.

One housekeeping note that catches people out. The old model names deepseek-chat and deepseek-reasoner were retired in July 2026. New projects should call the V4 model names directly.

Key Features

• Open weights. Download and self-host the models, which removes the data residency problem entirely for teams with hardware.
• One million token context. Both V4 models handle very large documents and codebases in a single request.
• Hybrid thinking mode. Switch step-by-step reasoning on or off within one model rather than choosing a different product.
• Drop-in API compatibility. OpenAI and Anthropic formats both supported, so existing code usually works with minimal changes.
• Context caching. Repeated prompt prefixes are billed at a fraction of the normal input rate, which suits chat and agent loops.
• Agent and coding strength. The July 2026 Flash update targets terminal, repository, and tool-use tasks specifically.
• Free consumer app. Web, Android, and iOS access at no cost, including reasoning and search.

Performance And Experience

On reasoning, maths, and code, DeepSeek performs at or near the level of models costing many times more. That is the headline, and it holds up. For long document analysis, the million-token window means fewer workarounds than most rivals require.

Writing quality is where the gap shows. English prose is accurate but plainer than the strongest Western models, and it needs more editing for anything published. Occasional Chinese characters appearing mid-answer were a known issue in earlier versions and have improved but not fully disappeared.

The free app is fast most of the time and slow when demand spikes, which is the expected trade for zero cost. Thinking mode always takes longer, since the model is working through the problem before replying.

Content handling reflects Chinese regulation. Politically sensitive topics, particularly those concerning China, are declined or answered narrowly in the hosted app. Self-hosted weights behave differently, which is one more reason technical teams prefer running them locally.

Pricing And The Cost Warning

The consumer app is free. API rates below are taken from DeepSeek’s official pricing documentation on 8 August 2026, per one million tokens.

For scale, mainstream Western frontier models commonly sit between five and thirty times higher on output tokens. That gap is the entire commercial argument for DeepSeek.

Read the footnote on DeepSeek’s own pricing page, though. The company states that it plans to raise API pricing in the near future and describes the expected increase as significant. If you are building a cost model on today’s rates, build in headroom.

Availability And Restrictions

This is the part most reviews get wrong, so it is worth stating precisely. There is no nationwide ban on ordinary consumers using DeepSeek in the United States. What exists is a patchwork of device-level restrictions.

Several US federal bodies, including the Navy, NASA, and parts of the Commerce Department, block DeepSeek on government equipment. States including Texas and New York have applied similar rules to their own devices and networks. Australia, Taiwan, South Korea, and India have restricted it for government use. Italy went furthest, with its data protection regulator ordering the consumer app removed from stores in early 2025.

The common thread is data location. DeepSeek’s privacy policy acknowledges storing user data on servers in China, and Chinese law grants authorities broad access to data held by domestic companies. Whether that matters depends entirely on what you are typing in.

Our Honest Verdict On DeepSeek

DeepSeek did something the industry needed. It proved that frontier-level capability does not require frontier-level pricing, and it published the weights so the claim could be checked. For a developer building on a budget, or a technical team that can run models on its own hardware, nothing else offers this combination. The V4 generation with its million-token window and hybrid thinking mode is a serious piece of engineering.

The caveats are not about capability. Data goes to China on the hosted service, and no amount of benchmark performance changes what that means for regulated or confidential work. Support is minimal. Content restrictions on political topics are real in the hosted app. And the company has now signalled a significant price rise, which weakens the one argument everyone leads with. Use the open weights if you can run them. Use the hosted app freely for personal work, learning, and code. Keep sensitive material out of it.

Suggestions For Improvement

• Offer a hosted option with data residency outside China for customers who need it.
• Publish the planned price increase with dates, so teams can plan rather than guess.
• Provide a clear enterprise support tier with response commitments.
• State content restrictions openly in the documentation instead of leaving users to discover them.
• Improve English writing quality, which trails the technical capability by a visible margin.
• Publish a plain-language data handling summary alongside the privacy policy.