Patchwire

DeepSeek · Released Sep 29, 2025

DeepSeek V3.2 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...

ReasoningOpen weights
  • Top 27%Fiction.LiveBench
  • Top 28%Aider Polyglot
Compare in showdown → OpenRouter ↗ All DeepSeek models
Released
Sep 29 2025
Context window
164K 147K out
Input / 1M tokens
$0.27
Output / 1M tokens
$0.41

Where to run it

Providers

Only SiliconFlow serves DeepSeek V3.2 Exp on OpenRouter right now.

ProviderInputOutputContextMax outputUptimevs. best
SiliconFlowFP8 $0.27 $0.41 164K147K99.9%24h Cheapest

Value

Price vs. performance

2 models score higher for the same price or lesson Terminal-Bench, e.g. GPT-5.1-Codex-Mini.
3.4%30.5%57.6%84.7% $0.01$0.10$1.00$10.00$100.00 Qwen3.6 35B A3B (Qwen) — 24.6% · $0.10/1M in GPT-5.5 (OpenAI) — 84.7% · $5.00/1M in Claude Opus 4.7 (Anthropic) — 80.2% · $5.00/1M in Grok 4.20 (xAI) — 57.3% · $1.25/1M in MiniMax M2.7 (MiniMax) — 45.1% · $0.30/1M in Qwen3.5-9B (Qwen) — 9.2% · $0.10/1M in GPT-5.4 (OpenAI) — 81.8% · $2.50/1M in GPT-5.3-Codex (OpenAI) — 78.4% · $1.75/1M in Gemini 3.1 Pro Preview (Google) — 80.2% · $2.00/1M in Claude Sonnet 4.6 (Anthropic) — 53.4% · $3.00/1M in MiniMax M2.5 (MiniMax) — 42.7% · $0.27/1M in GLM 5 (Z.ai) — 52.4% · $0.60/1M in Claude Opus 4.6 (Anthropic) — 79.8% · $5.00/1M in Kimi K2.5 (Moonshot AI) — 43.2% · $0.50/1M in GPT-5.2-Codex (OpenAI) — 66.5% · $1.75/1M in MiniMax M2.1 (MiniMax) — 36.6% · $0.30/1M in GLM 4.7 (Z.ai) — 33.4% · $0.54/1M in Gemini 3 Flash Preview (Google) — 64.3% · $0.50/1M in GPT-5.2 (OpenAI) — 64.9% · $1.75/1M in GPT-5.1-Codex-Max (OpenAI) — 60.5% · $1.25/1M in Claude Opus 4.5 (Anthropic) — 63.1% · $5.00/1M in GPT-5.1 (OpenAI) — 47.6% · $1.25/1M in GPT-5.1-Codex (OpenAI) — 57.8% · $1.25/1M in GPT-5.1-Codex-Mini (OpenAI) — 61.6% · $0.25/1M in Kimi K2 Thinking (Moonshot AI) — 35.7% · $0.60/1M in MiniMax M2 (MiniMax) — 30.0% · $0.26/1M in Claude Haiku 4.5 (Anthropic) — 35.5% · $1.00/1M in GLM 4.6 (Z.ai) — 24.5% · $0.55/1M in Claude Sonnet 4.5 (Anthropic) — 46.5% · $3.00/1M in GPT-5 (OpenAI) — 49.6% · $1.25/1M in GPT-5 Mini (OpenAI) — 34.8% · $0.25/1M in GPT-5 Nano (OpenAI) — 21.8% · $0.05/1M in gpt-oss-120b (OpenAI) — 18.7% · $0.15/1M in gpt-oss-20b (OpenAI) — 3.4% · $0.018/1M in Claude Opus 4.1 (Anthropic) — 38.0% · $15.00/1M in Gemini 2.5 Flash (Google) — 17.1% · $0.30/1M in Gemini 2.5 Pro (Google) — 32.6% · $1.25/1M in Gemini 2.5 Pro Preview 06-05 (Google) — 32.6% · $1.25/1M in DeepSeek V3.2 (DeepSeek) — 39.6% · $0.28/1M in DeepSeek V3.2 Exp (DeepSeek) — 39.6% · $0.27/1M in DeepSeek V3.2 Exp

DeepSeek V3.2 Exp Other DeepSeek Other labs Value frontier · Terminal-Bench vs. input price per 1M tokens (log scale) · Drag to pan, pinch or Ctrl + scroll to zoom, double-click to zoom in.

Benchmarks

11 benchmark scores

BenchmarkScoreRankLeader
Terminal-BenchSoftware engineering · Tasks completed with the programs in a real terminal (best agent) 39.6% #23of 40 GPT-5.584.7%
SciCodeSoftware engineering · Research-coding problems from the natural sciences 38.9% #77of 93 Claude Opus 5.566.9%
Aider PolyglotSoftware engineering · Hard Exercism problems in many languages 74.2% #3of 11 GPT-588.0%
APEX-AgentsAgents & computer use · Long professional tasks in banking, consulting and law 21.3% #49of 51 Gemini 4 Argon82.4%
Vending-Bench 2Agents & computer use · Run a simulated vending business for a year (final balance) $1,034 #47of 61 GPT-6 Astra$15,515
Fiction.LiveBenchWriting & content · Understanding long creative writing (16k tokens) 83.3% #4of 15 GPT-597.2%
WebDev ArenaDesign & visual · People vote on which model builds the better website (Elo) 1367 #72of 88 Claude Opus 5.51820
ARC-AGI-2Reasoning & knowledge · Few-shot abstract reasoning on grid puzzles 4.0% #59of 63 GPT-6 Astra95.0%
ARC-AGI-1Reasoning & knowledge · Pattern induction from input–output examples 57.0% #53of 63 Gemini 4 Argon98.5%
CritPtReasoning & knowledge · Open-ended, research-level physics problems 2.9% #65of 96 GPT-5.6 Sol32.3%
LMCAReasoning & knowledge · Judge the quality of conceptual arguments like experts 29.1% #81of 96 Claude Sonnet 5.568.3%

Specs

Details

Model ID
deepseek/deepseek-v3.2-exp
Developer
DeepSeek
Release date
Sep 29, 2025 ·
Context window
163,840 tokens
Max output
147,456 tokens
API price
$0.27 / $0.41 per 1M input / output tokens
Providers
1
Input types
text
Reasoning
Yes
Tool calling
Yes
Open weights
Yes
Free tier
No

Release date, prices and providers from OpenRouter's catalog; buying directly from a provider can cost a little differently. Uptime is the share of OpenRouter requests each provider served over the last day. Benchmark scores are the best result per benchmark from Epoch AI, licensed CC BY 4.0. Rank counts every model tracked here with a score.