DeepSeek V3.2 API

DeepSeek's DeepSeek V3.2, through an OpenAI-compatible API. Providers quote it on an open exchange, and each request goes to the offer with the lowest estimated charge that meets your rules.

Providers compete for each request. The exchange ranks the offers that meet your rules by estimated charge. It fixes your maximum charge before the first token, and you pay for the tokens that the model uses.
The OpenAI and Anthropic APIs. Use the OpenAI or Anthropic SDK that you have. Set the base URL and your key, and send deepseek-v3.2 as the model.
Your data rules. Send x-liquid-retention: none and only offers whose provider does not keep your content compete for the request.

Facts

Model name: deepseek-v3.2
Maker: DeepSeek
Context window: 160K tokens
Input price: $0.325 per million tokens
Output price: $0.475 per million tokens
Providers: 1 independent provider
Regions: us

Providers

Pera: us, $0.325 input and $0.475 output per million tokens, $0.325 cache read per million tokens, 160K tokens of context

Scores

HLE: 11.2, #151 of 273
GPQA: 75.1, #130 of 250
AA-LCR: 45.7, #169 of 237
Terminal-Bench Hard: 32.6, #49 of 166
LiveCodeBench: 59.3, #45 of 130
AIME 2025: 59.0, #41 of 100
MMLU-Pro: 83.7, #24 of 133
τ²-Bench: 78.9, #66 of 169
IFBench: 49.0, #90 of 178
SimpleQA Verified: 47.0, #33 of 126

Scores: Artificial Analysis (https://artificialanalysis.ai/)

Scores: Epoch AI, 'Capabilities & Benchmarking' (https://epoch.ai/benchmarks)

Request

curl https://router.inference.ai.exchange/v1/chat/completions \
  -H "Authorization: Bearer $LIQUID_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v3.2",
    "messages": [{"role":"user","content":"Say hello."}]
  }'

Questions

What is DeepSeek V3.2?: Served by intent-router: each job goes to the cheapest healthy upstream for this model.
How much does DeepSeek V3.2 cost?: Offers now ask $0.325 per million input tokens and $0.475 per million output tokens. A job pays for the tokens it uses. The exchange picks the cheapest eligible offer.
What is the context length of DeepSeek V3.2?: The largest offer takes 160,000 input tokens. One job can generate up to 65,536 tokens. A smaller offer can take less. The exchange skips an offer that cannot take your message.
What can DeepSeek V3.2 do?: Its providers list Reasoning effort, JSON object, Frequency penalty, Presence penalty, Seed, Stop sequences, Top-p sampling, Text, Tools.
What inputs and outputs does DeepSeek V3.2 support?: Text → Text.
Which providers serve DeepSeek V3.2?: 1 independent provider quote it now, with 1 offer. The table above ranks them by the cost of a typical job.
How do I call DeepSeek V3.2?: Send "deepseek-v3.2" as the model in an OpenAI or Anthropic request. The API section links to setup steps for each client.
When was DeepSeek V3.2 listed?: The exchange first listed it on 10/8/2026. This is not its release date.