DeepSeek is the Chinese AI lab that reset expectations about how much compute a frontier model actually requires. Founded by Liang Wenfeng inside the High-Flyer quantitative hedge fund, the lab has shipped a series of models — V2, V3, R1, V4 — whose efficiency and openness have made it one of the most consequential AI companies in the world, despite a small team and a fraction of US lab budgets.
This page is the living reference to DeepSeek — the model family, what it means for the broader AI landscape, pricing, how to use it, and every story we publish tagged DeepSeek. We update it whenever DeepSeek ships a meaningful change.
As of April 2026, DeepSeek-V4 is the current flagship, released under an MIT license with a 2M-token context window and native multimodal capabilities.
We rated DeepSeek 4.4 / 5. Genuinely impressive open-source AI at a price that breaks the rest of the market. If you're cost-sensitive or building on the API, DeepSeek deserves a serious look. The chat product is solid; the API is the real story.
Read the full DeepSeek review →Latest news about DeepSeek
Every story we've published tagged DeepSeek. Refreshed automatically every time the newsroom publishes.
- Aug 22, 2026Ulanqab emerges as China's biggest AI data center hub with 12.5GW pledgedDeepSeek, ByteDance and Alibaba are among the Chinese firms racing to build in the Inner Mongolian city, dwarfing OpenAI's 10GW Stargate footprint.
- Aug 15, 2026Anthropic Q2 revenue hits $11.5B, a 14-fold jump ahead of IPOClaude's maker crossed $11.5B in a single quarter and posted positive adjusted operating income as it lines up a fall listing.
- Aug 14, 2026OpenAI cuts GPT-5.6 Luna 80% as Anthropic undercuts its own flagshipUS token prices have dropped nearly a quarter since mid-July as DoorDash and Airbnb shift workloads to Chinese models from Moonshot and DeepSeek.
- Jul 23, 2026DeepSeek founder says AGI trumps profit, top models to stay open-sourceLiang Wenfeng told Yicai that DeepSeek will keep releasing its frontier models as open weights, prioritising AGI over commercial returns.
- Jul 20, 2026UK AISI: open-weight models now trail closed AI on cyber by 4-7 monthsGLM-5.2 and DeepSeek V4-Pro closed the cyber capability gap from 6-10 months to 4-7 months. Kimi K3 is next in line.
- Jul 20, 2026LLMs stereotype job candidates 65% more than humans in Princeton hiring studyOpenAI's o3 scored 1.83 on a segregation scale where 2 is total sorting — and reasoning models were the worst offenders.
Timeline of major releases
- Jul 17, 2023DeepSeek foundedDeepSeek spins out of High-Flyer Quant, a Chinese hedge fund, with a mandate to build frontier AI. Founder Liang Wenfeng funds it from the fund's GPU stockpile.
- May 7, 2024DeepSeek-V2 and the pricing disruptionDeepSeek-V2 launches with API pricing roughly 1/10 of GPT-4. The model is strong enough on benchmarks to force Chinese competitors into an AI price war.
- Dec 26, 2024DeepSeek-V3 shipsDeepSeek-V3 (671B MoE) is released with a paper disclosing training cost of ~$5.6M. The efficiency claim kicks off global debate about compute moats.
- Jan 20, 2025DeepSeek R1 and the January 27 market shockDeepSeek R1 launches as a chain-of-thought reasoning model competitive with OpenAI's o1. On Jan 27, 2025, NVIDIA drops 17% and global tech stocks lose ~$1T on efficiency concerns.
- Aug 21, 2025DeepSeek V3.1 and Anthropic API compatibilityV3.1 ships with an Anthropic-compatible API format, making DeepSeek a drop-in replacement for Claude in coding tools. Developer adoption in the West accelerates.
- Mar 10, 2026DeepSeek-V4 launchesDeepSeek-V4 ships with native multimodal, extended reasoning, and a 2M context window. Continues the pattern of releasing open weights and publishing detailed technical reports.
How to access DeepSeek
| Plan | Price | Includes |
|---|---|---|
| Free (chat.deepseek.com) | $0 | DeepSeek-V4 via web chat and mobile app, with rate limits. |
| DeepSeek API | Usage-based (very low) | DeepSeek-V4 via API. Pricing is among the lowest for a frontier model — roughly 1/10 of Claude or GPT-5 per token. |
| Weights | Free | Model weights released openly on Hugging Face under the MIT License (for V3/V4). |
| Third-party hosted | Usage-based | DeepSeek available on most inference providers — Together, Fireworks, Groq, plus Chinese clouds. |
DeepSeek vs. ChatGPT vs. Claude
| Feature | DeepSeek | ChatGPT | Claude |
|---|---|---|---|
| License model | Open weights (MIT) | Closed / hosted only | Closed / hosted only |
| API pricing (rough) | ~$0.30–2 per 1M tokens | ~$5–15 per 1M tokens | ~$3–15 per 1M tokens |
| Reasoning benchmarks | Competitive with o3 and Claude Sonnet | Frontier (GPT-5 Pro) | Frontier (Opus 4.7) |
| Self-hostable | Yes (weights released) | No | No |
| Origin | China | USA | USA |
The DeepSeek efficiency story
DeepSeek's central contribution to the AI field isn't a specific model — it's a collection of training and architecture choices that compound into dramatically lower training costs. Mixture-of-experts routing, reinforcement learning from reasoning traces, mixed-precision training, aggressive dataset curation: individually, these are known techniques. Together, as DeepSeek assembled them, they produced GPT-4-class reasoning at a fraction of the publicly disclosed budgets of US frontier labs.
The $5.6M training cost figure from the V3 paper — even with all the caveats about what that figure does and doesn't include — reframed the "compute moat" conversation that had dominated AI investor thinking in 2024. It's a major reason the 2026 AI landscape has more credible frontier players than it otherwise would.
Using DeepSeek in 2026
There are three ways to use DeepSeek today. The consumer chat app at chat.deepseek.com is free and works like ChatGPT. The DeepSeek API is among the lowest-priced frontier APIs available — we routinely see teams use it for high-volume workloads where Claude or GPT-5 per-token costs would be prohibitive. And the open weights (released on Hugging Face under MIT) let you self-host or use via third-party inference providers like Together or Groq.
For most Western developers, the V3.1 release was the turning point — it ships an Anthropic-compatible API format, so you can point your existing Claude Code or Anthropic SDK at DeepSeek endpoints and it just works. That ergonomic compatibility, combined with the pricing gap, is why DeepSeek has shown up in a lot of "cost-first" AI deployments over the past year.
Data and privacy considerations
DeepSeek is operated from China, and data sent to the hosted chat or API may be subject to Chinese data laws. For privacy-sensitive use cases — regulated industries, proprietary code, customer data — the right call is usually to run the open weights on your own infrastructure or on a Western-hosted provider that serves DeepSeek models.
The open-weights release is the genuine advantage here. Unlike Western closed labs, DeepSeek gives you the option to sidestep the hosting question entirely.
Where DeepSeek wins — and where it doesn't
DeepSeek wins on price-to-capability ratio, openness (true MIT-licensed weights), and reasoning-specific benchmarks. For high-volume coding, math, and reasoning workloads where cost matters, it's now a default option.
It trails on polish (the chat product is less refined than ChatGPT or Claude), ecosystem (fewer third-party integrations than the major Western labs), and trust (Chinese-origin services carry compliance friction in many enterprise contexts). For most consumer use, ChatGPT or Claude remain the better pick; for developer workloads where cost matters, DeepSeek is the obvious contender.
Our coverage
Everything tagged DeepSeek appears in Latest news above. For broader coverage of the US-China AI competition, export controls, and compute moats, see our reporting under the AI news and AI business beats. For other open-weights alternatives, see Llama and Mistral.
