DeepSeek V3.2 Thinking

DeepSeek V3.2 (thinking/reasoner mode) — official successor to V3.2-Exp. Reasoning-first model built for agents with GPT-5 level performance. Balanced inference vs. output length. First DeepSeek model with thinking-in-tool-use capability. FP8.

Reasoning
Native PDF input
Tool Calling
Structured Output

Added Dec 1, 2025

Model weights

Pricing

Auto routing · per 1M tokens

Input: $0.21
Output: $0.31
Cache read: $0.022

Compare provider prices

Specifications

Context window: 163K
Max output: 65.5K
Parameters: 685B / 37B; Total / active
Avg output (7d): 1.2K tokens; Longer than 78% of models

Benchmarks

Artificial Analysis

LMArena

Vectara

Sourced from Artificial Analysis.

Intelligence Index

21.5

Better than 68% of models compared

Coding Index

44.2

Better than 52% of models compared

Agentic work

T²-Bench Telecom (legacy)

Legacy fallback · Conversational AI agents in dual-control scenarios

90.6%

Better than 85% of models compared

GDPval-AA v2

Economically valuable tasks (Elo)

695 Elo

Better than 32% of models compared

Document reasoning

AA-LCR v1.1

Long context reasoning with updated grading

73.3%

Better than 69% of models compared

Reasoning

HLE

Humanity's Last Exam

24.6%

Better than 73% of models compared

IFBench

Instruction-following benchmark

60.7%

Better than 72% of models compared

CritPt

Research-level physics reasoning

2.9%

Coding

Terminal-Bench Hard (legacy)

Legacy fallback · Agentic coding and terminal use

35.6%

Better than 82% of models compared

LiveCodeBench

Contamination-free coding benchmark

86.2%

Better than 97% of models compared

Math

AIME 2025

American Invitational Mathematics Examination 2025

92.0%

Better than 94% of models compared

Knowledge

MMLU-Pro

Professional and academic subject knowledge

86.2%

Better than 95% of models compared

AA-Omniscience Accuracy

Proportion of correctly answered questions

33.0%

AA-Omniscience Hallucination Rate

Rate of incorrect answers among non-correct responses

82.7%

Legacy benchmarks

GPQA Diamond (legacy)

Graduate-level scientific reasoning

84.0%

Better than 75% of models compared

AA-LCR (unversioned / legacy)

Long context reasoning evaluation

73.3%

Better than 69% of models compared

GDPval-AA (unversioned / legacy)

Economically valuable tasks

9.0%

Last updated Oct 3, 2026