DeepSeek V3.2 Exp

ModelOpen sourceActive

by DeepSeek · family “deepseek-v3-exp”

DeepSeek-V3.2-Exp is an experimental large language model released by DeepSeek as an intermediate step between V3.1 and future architectures. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism... (description from the OpenRouter listing)

in: textout: textReasoningTool useStructured output

Epoch Capabilities Index

145.0 #81 of 268

90% CI 142.3 – 146.9

Listing ↗

Released

29 Sep 2025

Context window

164K tokens

Max output

66K tokens

Input $ / 1M tokens

$0.27

Output $ / 1M tokens

$0.41

Cached input $ / 1M

Unknown

Each benchmark shown separately with its own source

Benchmark results (8)

All benchmarks

Benchmark	Domain	Score	vs best recorded	Setting	Run	Source
LMCA	agents	29.1%	44%	thinking	—	External ↗
DTBench	reasoning	87.7%	89%	thinking	—	External ↗
CL-bench Life	long-context	9.5%	43%	thinking	—	External ↗
CL-bench	long-context	13.2%	47%	thinking	—	External ↗
Fiction.LiveBench	long-context	83.3%	86%	high	—	External ↗
WeirdML	coding	39.5% ±0.0	42%	thinking	—	External ↗
Aider polyglot	coding	74.2%	84%	thinking	—	External ↗
The Agent Company	agents	42.9%	100%	—	—	External ↗

Source: Epoch AI Benchmarking Hub (CC BY 4.0). “External” rows are leaderboard results Epoch collects from third parties. Best reported setting per benchmark is shown.

API list price over time

Price history

$ per 1M tokens

13 observations (OpenRouter listing + Internet Archive snapshots). History accumulates with every ingest run; a single point means no change has been observed yet.