Pass IndexThe State of AISign in

DeepSeek V3.2 Exp

DeepSeek-V3.2-Exp is an intermediate step toward the next-generation architecture of the DeepSeek models by introducing DeepSeek Sparse Attention—a sparse attention mechanism designed to explore and validate optimizations for training and inference efficiency in long-context scenarios.

text → text · made by DeepSeek

$0.21$0.32per Mtok in / out
DeepInfra · 8 sellers

Sold by 9 ways

SellerLaneRate
DeepInfra aggregatorstandard$0.21per Mtok in$0.32per Mtok outapi.deepinfra.com · read 2026-08-25
ElectronHub aggregatorstandard$0.27per Mtok in$0.41per Mtok outapi.electronhub.ai · read 2026-09-16
Hugging Face Inference Providers aggregatornovita$0.27per Mtok in$0.41per Mtok outrouter.huggingface.co · read 2026-08-25
Nous Research aggregatorstandard$0.27per Mtok in$0.41per Mtok outinference-api.nousresearch.com · read 2026-09-12
Novita AI aggregatorstandard$0.27per Mtok in$0.41per Mtok outapi.novita.ai · read 2026-08-25
OpenRouter 2 ways$0.27per Mtok in$0.41per Mtok out
OpenRouter aggregatorstandard$0.27per Mtok in$0.41per Mtok outopenrouter.ai · read 2026-08-24
OpenRouter aggregatoratlas-cloud/fp8$0.27per Mtok in$0.41per Mtok out$0.27per Mtok cachedopenrouter.ai · read 2026-08-25
SiliconFlow aggregatorstandard$0.27per Mtok in$0.41per Mtok outmodels.dev · read 2026-09-16
PPIO aggregatorstandard$2per Mtok in$3per Mtok outapi.ppinfra.com · read 2026-09-16

Measured 19 standings

PlaceBoardMetricScore
1stof 16The agent company — Epoch AI% Score0.524
12thof 69Aider polyglot coding leaderboardPercent correct74.2
14thof 109Berkeley Function-Calling Leaderboard (BFCL) V4Overall Acc (prompt + thinking)56.73
15thof 61AIME 2025Accuracy (± 95% CI) (think)91.67
16thof 105Vectara Hallucination LeaderboardHallucination Rate5.3
19thof 109Berkeley Function-Calling Leaderboard (BFCL) V4Overall Acc (fc)54.12
21stof 72Aider polyglot — Epoch AIPercent correct70.2
46thof 60Vending bench 2 — Epoch AIScore1034.004
79thof 147Design Arena · 3d (models)Elo1176
83rdof 150Design Arena · uicomponent (models)Elo1182
85thof 156Design Arena · codecategories (models)Elo1186
89thof 162Design Arena · website (models)Elo1191
90thof 155Design Arena · gamedev (models)Elo1167
95thof 109Design Arena · svg (models)Elo1057
98thof 152Design Arena · dataviz (models)Elo1168
108thof 121Webdev arena — Epoch AIArena Score1271.77
111thof 128LMArena · WebDevScore (Elo)1272
116thof 402LMArena · TextScore (Elo)1422
335thof 656Epoch capabilities index — Epoch AIECI Score145.06

About

DeepSeek V3.2 Exp — a text model from DeepSeek, sold by 8 companies from $0.21 in and $0.32 out per million tokens, placed 1st of 16 on The agent company — Epoch AI.

It takes text and returns text, with a context window of 163,840 tokens. It was published in September 2025, trained on material up to July 2025. Its sellers say it can call a tool. The catalogue files it under extract, reasoning, code and agents. Eight companies sell it. The cheapest is $0.21 in and $0.32 out per million tokens at DeepInfra. Beside the standard rate there are separately routed and regional lanes. It has been measured on 18 boards, and stands best at 1st of 16 on The agent company — Epoch AI.

Every current figure

Maker
DeepSeek
Register
model
Takes
text
Returns
text
Context
163,840 tokens
Published
September 2025
Knowledge to
July 2025
Parameters
685 billion
Licence
not read
Sellers
8
Maker's own price
not read
Price
$0.21 in and $0.32 out per million tokens — DeepInfra
Boards
18
Best place
1st of 16 — The agent company — Epoch AI

Known as 7 names

DeepSeek-V3.2-Exp (Chat)DeepSeek-V3.2-Exp (Reasoner)DeepSeek-v3.2-Exp (Think)DeepSeek-V3.2-Exp (FC)DeepSeek-V3.2-Exp (Prompt + Thinking)deepseek-ai/DeepSeek-V3.2-Expdeepseek/deepseek-v3.2-exp