DeepSeek
Hangzhou lab shipping MIT-licensed V4 models and a very cheap API
Overview
DeepSeek is a Hangzhou AI lab that publishes open-weight models under the MIT licence and runs a free chat app and a low-cost API. Its current models are V4-Pro (1.6 trillion parameters, 49 billion active) and V4.1-Flash, released on 10 September 2026 with native image understanding, both with a 1M-token context. API prices halve outside weekday peak hours. (See also our dedicated DeepSeek V4 entry.)
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- Expert Mode and Instant Mode
- Deep thinking
- Web search
- DeepSeek-V4.1-Flash, 1M context
- Vision supported
- Cache-hit input $0.006 per 1M at peak
- 2,500 concurrent requests
- DeepSeek-V4-Pro-0813, 1M context
- Thinking and non-thinking modes
- Cache-hit input $0.044 per 1M at peak
- 500 concurrent requests
- V4-Pro, V4-Flash and V4.1-Flash weights
- Self-hosting on your own hardware
API pricing (pay as you go)
| Model | Input | Cached input | Output |
|---|---|---|---|
| deepseek-flash (DeepSeek-V4.1-Flash) Off-peak: $0.15 miss / $0.003 hit / $0.60 out. Legacy names deepseek-v4-flash* are now served by V4.1-Flash. | $0.30 | $0.006 | $1.20 |
| deepseek-v4-pro (DeepSeek-V4-Pro-0813) Off-peak: $0.66 miss / $0.022 hit / $1.98 out. Service continues past 14 Sep 2026 at unchanged billing. | $1.32 | $0.044 | $3.96 |
Use cases
What you can produce with DeepSeek
- Free chat with deep thinking and web search on the web and in mobile apps
- An OpenAI- or Anthropic-format API with 1M-token context from $0.15 per million input tokens off-peak
- Image understanding through the V4.1-Flash API
- Self-hosted, air-gapped inference from MIT-licensed V4 weights
- Coding-agent sessions in Claude Code, OpenCode or Codex pointed at the DeepSeek API
- Proof-style mathematics from the separate DeepSeek-Math-V2 model, which DeepSeek reports at gold level on IMO 2025
DeepSeek remains the price benchmark for capable models: V4.1-Flash costs $0.15 to $0.30 per million input tokens and V4-Pro under $4 per million output at peak, both with 1M-token context, and the weights are MIT-licensed for self-hosting. The free app is popular and well rated.
The hosted service stores data in China, several governments restricted it in 2025, and the models carry Chinese content moderation. Regulated ASEAN teams should self-host; others should schedule bulk jobs outside Singapore's 09:00-12:00 and 14:00-18:00 weekday peak, when rates halve.
What we found
DeepSeek (深度求索) is the Hangzhou AI lab whose open-weight V3 and R1 models, released in December 2024 and January 2025, upset assumptions about what frontier-class AI costs to build. Its current line is V4. A preview on 24 April 2026 open-sourced V4-Pro, a mixture-of-experts model with 1.6 trillion parameters and 49 billion active, and V4-Flash at 284 billion and 13 billion, both with 1M-token context and MIT licences. V4-Pro reached general release on 13 August with stronger agent scores, and on 10 September V4.1-Flash replaced V4-Flash: a 552-billion-parameter model with native image understanding that DeepSeek says beats V4-Pro on its own benchmarks. Benchmark figures here are DeepSeek's. The lab also offers a desktop agent, DeepSeek Harness, and specialised releases such as DeepSeek-Math-V2.
The API accepts OpenAI and Anthropic formats. Since 16 August prices have had peak and off-peak rates. deepseek-flash costs $0.30 per million input tokens on a cache miss at peak and $0.15 off-peak, with output at $1.20 or $0.60; V4-Pro is $1.32 in and $3.96 out at peak. Peak means 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which is 09:00 to 12:00 and 14:00 to 18:00 in Singapore, so ASEAN office hours largely pay the full rate. Old model names are a trap: deepseek-chat and deepseek-reasoner were retired in July 2026, and the V4-Flash names now route to V4.1-Flash.
The chat app and website are free; on 14 October 2026 Google Play in Singapore showed 4.4 stars from 343,000 reviews. Data residency is the main caveat. DeepSeek's privacy policy says it collects, processes and stores personal data in the People's Republic of China. In early 2025 Italy's data protection authority ordered the app blocked, Australia barred it from government devices and South Korea's privacy regulator suspended new downloads. Regulated teams usually self-host the open weights instead, which removes the residency question but not the content moderation trained into the models.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- DeepSeek R1 was released under the MIT licence — the most permissive open-source licence possible — allowing anyone to use it commercially without restrictions.
- The reported training cost of DeepSeek V3 was approximately $5.6 million, compared to estimated hundreds of millions for comparable Western frontier models.
- DeepSeek's app briefly became the #1 downloaded app in the US App Store in January 2025, triggering a 17% drop in NVIDIA's stock price.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including DeepSeek's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with DeepSeek unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to DeepSeek directly →
Spotted something out of date? Suggest an update →
DeepSeek: pricing and development
The cheapest frontier tier, tracked as it moves — endpoint retirements, benchmark claims, and the shift to charging by the clock.
- DeepSeek Starts Charging Agents for Web Search. Claude Code and Codex Users Pay From 1 November
- DeepSeek's Funding Round Doubled to 100 Billion Yuan. Its Valuation Did Not Move
- DeepSeek Planned to Answer V4 Pro Calls With a Different Model. Then It Backed Down.
- DeepSeek Hired Bankers. That Is Not the Same as Filing to List.
DeepSeek in the news
From our guides
Guide · 9 min read
Why It Confidently Makes Things Up
We asked two AI models for five peer-reviewed studies on an obscure topic, with...
Guide · 16 min read
Self-Host Dify with Docker Compose: 15 Containers to Your Own LLM App Studio (2026)
We deployed Dify 1.16.0 on Docker Desktop, hit a real port clash, wired DeepSeek...
Guide · 9 min read
The Best Open-Source LLMs You Can Actually Self-Host (2026)
Qwen, Llama, DeepSeek, Mistral, Phi and Gemma compared for self-hosting in 2026...
Alternatives to DeepSeek
More in LLMs & Chat
