DeepSeek · open · 2026-04
DeepSeek-V4-Pro
#3 index 88.1
the takeThe open-weights model that actually reaches the frontier, at prices that should embarrass the closed labs. The asterisk is loud: these are vendor numbers in max-reasoning mode, not yet independently replicated. If they hold, it is the story of the year.
SWE-bench Verified
80.6
GPQA Diamond
90.1
MMLU-Pro
87.5
AIME (math)
95.2
- Context
- 1M
- Input
- $0.435/M
- Output
- $0.87/M
- Speed
- 89.9 tok/s
- Modality
- text
- Frontier scores at a fraction of the price
- MIT-licensed, self-hostable, 1M context
- Strong agentic and competitive coding
- Headline scores vendor-reported, unverified
- 1.6T params make self-hosting heavy
- Text-only; top scores need expensive Think-Max