DeepSeek review, pricing and verdict

Last updated: 2026-11-02

DeepSeek is a Chinese AI lab's free chat app and per-token API. Its default model is a 552-billion-parameter Mixture-of-Experts design released in September 2026 that reads images as well as text and runs in thinking or non-thinking mode. The weights are published on Hugging Face under the MIT licence.

About DeepSeek

DeepSeek is an AI lab based in Hangzhou, China, founded in 2023. Its R1 reasoning model drew worldwide attention in early 2025, and the lab has since kept up a fast run of open-weight releases; the DeepSeek company page covers the organisation itself.

The current default is DeepSeek V4.1 Flash, released on 10 September 2026 and updated to the V4.1-Flash-1017 build on 17 October. DeepSeek describes it as a 552-billion-parameter Mixture-of-Experts model with a new encoder-decoder design, and it replaced V4 Flash as the first model in the line that understands images natively. DeepSeek V4 Pro, which reached general availability in August, stays on the API after DeepSeek dropped a plan to route its traffic to Flash, but it does not accept images. Both share the same context window and output cap, and both ship as MIT-licensed weights on Hugging Face.

The free web chat and the iOS and Android apps run the latest Flash build. Developers use a metered API that accepts both the OpenAI and the Anthropic request format, so agents such as Claude Code and OpenCode can point at it with a base-URL change, and the lab's own coding agent runs as a desktop app on Windows and macOS. In October DeepSeek opened its in-house search engine to developers as paid Search and Fetch APIs.

Who should look elsewhere: DeepSeek's site lists no team plan, admin console or paid tier for the chat app, so a company that needs SSO, audit logs or data kept outside China will find more admin tooling in ChatGPT, and buyers in that position should also look at Claude before deciding. For a fuller agent app on similar open-weight terms there is Kimi, and teams inside Alibaba Cloud can start with Qwen instead. If a licence matters more than price, Mistral Large 3 is an open-weight option under Apache 2.0. Prices also move with the clock, because peak-hour rates are double the off-peak ones, so a batch job scheduled at the wrong time costs twice as much. For a wider comparison, see the best AI chatbots guide.

Pricing

The DeepSeek web chat and mobile apps are free. API prices in US dollars per million tokens, peak / off-peak: deepseek-flash (build V4.1-Flash-1017) $0.30 / $0.15 input on a cache miss, $0.006 / $0.003 on a cache hit and $1.20 / $0.60 output; deepseek-v4-pro $1.32 / $0.66 input on a cache miss, $0.044 / $0.022 on a cache hit and $3.96 / $1.98 output. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays, excluding Chinese public holidays; every other hour bills at half.

The Search and Fetch APIs cost $0.75 per 1,000 successful requests each, and their use inside DeepSeek's desktop coding agent, Claude Code and Codex became billable on 1 November 2026. A granted promotional balance is spent before topped-up credit. Read on 2 November 2026 from DeepSeek's Models & Pricing page.

Key Features

• Flash as the default model: The September 2026 Flash release activates 8 billion parameters for input and 16 billion for output, per DeepSeek, and powers both the free app and the deepseek-flash API.
• Long context and output: Both API models read up to 1M tokens and can write up to 384K tokens in one response.
• Thinking and non-thinking modes: Thinking mode is on by default on both models and can be switched off per request; fill-in-the-middle completion works only with thinking off.
• Native image input on Flash: Image understanding arrived with the September Flash release; V4 Pro remains text-only on the API.
• OpenAI and Anthropic request formats: One base URL serves the OpenAI format and a second /anthropic path serves the Anthropic format, with JSON output, tool calls and a Responses API on both models.
• Search and Fetch APIs: Opened on 17 October 2026 on DeepSeek's in-house search engine and billed per 1,000 successful requests.
• MIT-licensed weights: Both current models can be downloaded from Hugging Face and carry the MIT licence, so they can be self-hosted or fine-tuned.
• Desktop coding agent: DeepSeek's own coding agent installs on Windows and macOS and works in the background and editing local files.

Pros

• Off-peak pricing gives a predictable way to halve API spend: schedule batch jobs outside the weekday peak windows.
• MIT-licensed weights let a team prototype on the hosted API and later self-host the same model with no licence negotiation.
• Accepting both OpenAI and Anthropic request formats means most existing agent tooling connects by changing a base URL and key.

Cons

• Peak windows fall in Asian business hours (09:00-12:00 and 14:00-18:00 in Singapore), so teams in that region pay the higher rate for daytime work.
• The lineup turns over quickly: V4 Flash was retired as soon as its successor shipped, and DeepSeek has said it is phasing out V4 Pro.
• V4 Pro does not accept images, so any vision work has to go to the Flash model.
• DeepSeek is a China-based company; teams with data-residency rules should read its privacy policy before sending regulated data.
• There is no paid consumer plan or team workspace for the chat app, so there is no admin control over how staff use it.

Visit DeepSeek Official Website