DeepSeek V4 Pro
Open 1.6T MoE with elite competitive programming at a tenth of closed-model prices.
Overall Intelligence
#27
of 56
10.5 pts behind #1
0
races led
25
weeks at #1
78.3
score
- Released
- Apr 24, 2026
- 6 mo ago
- Context
- 1M
- 384K max output
- Price
- $1.74 / $3.48
- per 1M in/out · $2.17 blended
- Speed
- 60 tok/s
- 5s to first token
- Modalities
- text
- outputs text
- Weights
- Open
- 1.6T MoE (49B active)
The 30-second version
An open 1.6-trillion-parameter mixture-of-experts model with elite competitive programming, priced at roughly a tenth of closed frontier models.
- 1Top of the open field on competitive programming and algorithms.
- 2$1.74 / $3.48 per million tokens: exceptional value.
- 3Open weights with a permissive licence.
- 4Less polished on agentic tool use and writing than the closed leaders.
What you need to know
Purpose
Frontier-adjacent reasoning and coding at a radically lower price, available to anyone.
What DeepSeek optimised for
Training efficiency and reasoning, especially math and algorithms, with open release as a strategy.
Unique strengths
- Algorithmic problem solving.
- Price.
- Open weights.
Where it ranks
Current snapshot · illustrative scores · click any race to watch it
V4 ProBest on each axis
Overall
Coding
Hard Reasoning
Science
Writing & Language
Multimodal
Tool Use & Agents
Speed & Cost
Overall Intelligence over time
Best use cases
01
Algorithmic coding
Contest-style problems, optimisation and data-structure work.
02
Cost-sensitive reasoning
Large batch jobs that need decent reasoning cheaply.
03
Research and fine-tuning
A strong open base for academic and enterprise customisation.
Strengths
Algorithms
Among the leaders of the competitive programming race.
Weaknesses
Agentic polish
Tool use is less reliable than the closed leaders.
When to choose it, and when not to
Pick something else when…
- Workloads with strict data-residency rules for its hosted API: self-host instead.
- Polished customer-facing writing.
Update history
The distilled version first; open any release for the deeper breakdown.
The DeepSeek V family