DeepSeek.
Open-weights distribution that reset the cost-per-token curve.
What it is.
Hangzhou-based research house that shipped V3 in late 2024 at a training cost an order of magnitude below GPT-4. R2 followed with frontier-comparable reasoning. Open weights under permissive licence. Reset what enterprise procurement teams believe a frontier-class model costs to train and serve.
Where it fits.
Cost-anchor in every frontier-model contract since 2025. Direct production use blocked at most US Fortune 500 by procurement policy on Chinese-origin software — but indirect pressure on pricing is visible across all American vendor contracts.
- Lowest cost-per-token at frontier-comparable quality
- Open weights enable any deployment topology
- Strong coding performance with DeepSeek Coder
- Chinese-origin restricts direct enterprise procurement in US and parts of EU
- API hosted in mainland China for non-self-hosted use
Frequently asked.
Can US Fortune 500 buyers use DeepSeek?
Indirectly — open weights served on US infrastructure are usually clearable. Direct API access to deepseek.com is blocked by most procurement policies.
Is DeepSeek really cheaper to train?
Verified by independent researchers within an order of magnitude. The architectural choices around MoE and inference are public and reproducible.
What is the practical reasoning quality?
R2 is within striking distance of GPT-5 and Claude Sonnet on public reasoning benchmarks. Below frontier on tool-use ecosystem and English long-tail.
