What does "winning" look like when you give your model away for free?
DeepSeek V4 benchmarks place it roughly on par with Claude Opus 4.6 - ahead of most models, but not at the level of GPT-5.5 or Opus 4.7. In head-to-head evals, the gap to the top closed models is real. In day-to-day practice, V4 performs somewhere around what practitioners describe as "GPT-5.2 level" - consistent, reliable, and useful across a wide range of tasks.
That framing - "it's not the best" - misses the actual story.
The Open-Source Math
DeepSeek releases its models as open weights, meaning anyone can download and run them on their own hardware. That changes the entire cost structure. Running V4 locally costs a fraction of what top-tier closed APIs charge per million tokens. For high-volume applications - customer support pipelines, batch document processing, code generation across an entire codebase - the economics are fundamentally different.
Being roughly equivalent to Opus 4.6 while running locally isn't a consolation prize. Opus 4.6 handles complex reasoning, long documents (V4's context window supports roughly 128,000 tokens, or about a 300-page book's worth of text), and nuanced writing tasks well. Getting that capability on self-hosted infrastructure is what the open-source community has been pushing toward for two years.
Benchmark Scores vs. Real Use
The gap between DeepSeek V4 and current frontier models is measurable on standard benchmarks - tasks like graduate-level reasoning problems, math competitions, and code generation. On those structured tests, V4 sits a clear tier below GPT-5.5 and Opus 4.7.
Real-world use narrows that gap. Benchmarks favor tasks with clean right-or-wrong answers. Most of what practitioners actually do - drafting, summarizing, answering questions about internal documents, writing code with business context - doesn't map neatly to benchmark categories. In those everyday tasks, the difference between "top tier" and "second tier" is often smaller than the scores suggest.
That doesn't mean V4 is secretly better than the benchmarks show. It means benchmarks and day-to-day utility are measuring different things.
The real question for anyone evaluating DeepSeek V4 isn't whether it can beat ChatGPT on standardized tests. It's whether the quality-to-cost ratio works for your specific situation. For organizations processing millions of tokens daily, or those that need to keep data on their own servers for compliance reasons, the answer is often yes - even if V4 never closes the gap to the absolute frontier.