DeepSeek

DeepSeek-V3.2

#3τ-bench

by DeepSeek

+
+
+
+
About

DeepSeek-V3.2, released by DeepSeek in December 2025, was the first model in the V3 series to integrate thinking directly into tool use — allowing the model to reason while calling external tools, rather than treating reasoning and action as separate stages. This made it particularly relevant for complex agentic workflows where tool calls need to be reasoned through rather than just executed, and it supports tool use in both thinking and non-thinking modes.

+
+
+
+
Pricing Range
Input (per 1M)$0.27 -$0.27
Output (per 1M)$1.10 -$1.10
Providers1
+
+
+
+
Timeline
ReleasedDec 1, 2025
+
+
+
+
License & Family
License
MIT
Performance Overview
Performance metrics and category breakdown

Overall Performance

2 benchmarks
Average Score
59.0%
Best Score
80.4%
High Performers (80%+)
1

Performance Metrics

Max Context Window
136.0K

Top Categories

Agents
80.4%
Coding
37.5%
+
+
+
+
All Benchmark Results for DeepSeek-V3.2
Complete list of benchmark scores with detailed information
τ-bench
Agents
80.40
80.4%
Unverified
SWE-rebench
Coding
37.50
37.5%
Unverified