DeepSeek V4 arrived in April with the same trick DeepSeek always pulls: numbers that should cost a fortune, at prices that do not. Top-tier coding scores, a million-token context, an MIT license, and output priced under a dollar per million tokens. The West still has not quite made peace with it.
The specs that matter
V4 comes in two sizes. V4-Pro is a 1.6-trillion-parameter mixture-of-experts model with 49 billion active. V4-Flash is a leaner 284 billion, 13 billion active. Both offer a one-million-token context, and both are MIT-licensed with weights on Hugging Face. Pro scores 80.6% on SWE-bench Verified, the highest open-weights result, level with Gemini 3.1 Pro, and posts a Codeforces rating that would place it among the better human competitive programmers.
Now the prices
V4-Pro runs about $0.44 per million input tokens and $0.87 per million output. Flash is a quarter of that. That is a fraction of what the closed frontier charges for comparable coding ability. DeepSeek made its steep discount the permanent list price, which is either confidence or a declaration of war on everyone margins, depending on where you sit.
When a model this capable costs cents, the question stops being "can we afford AI" and becomes "why are we still paying premium prices for the easy 90 percent."
The uncomfortable part
A lot of the discomfort around DeepSeek is not technical. It is that a Chinese lab keeps shipping open, permissively-licensed models that embarrass the pricing of Western incumbents, and people would rather find a reason to dismiss it than update. The honest position is simpler: the weights are public, the license is MIT, the benchmarks are reproducible, and the code it writes is good. You do not have to love the geopolitics to use the tool well.
Where it fits
- Bulk coding and automation where you would otherwise burn money on a frontier API.
- Self-hosting, if you have the hardware. Pro is an 865GB download, Flash a more manageable 160GB.
- A cheap second opinion to cross-check a more expensive model, a habit more teams should build.
Not the most capable model in the world. Arguably the most capable per dollar, and in most real work that is the number that decides the budget. Quietly excellent is exactly right.