DeepSeekReasoningReleased 1 December 2025
DeepSeek V3.2 pricing, context and benchmark
DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance.
164K
Input context
147K
Output context
$0.209
Official input
$0.31
Official output
n/a
AA index
01 · Model Overview
Everything that matters, in one table
Context, pricing, cache, modalities and tooling, from the model registry and Artificial Analysis.
| Model | DeepSeek V3.2 |
| Model ID | deepseek/deepseek-v3.2 |
| Provider | DeepSeek |
| Release date | 1 December 2025 |
| Status | Generally available |
| Input context | 163,840 tokens |
| Output context | 147,456 tokens |
| Official input pricing | $0.209 per 1M tokens |
| Official output pricing | $0.31 per 1M tokens |
| Cache read pricing | $0.13 per 1M tokens |
| Market input pricing | $0.209 per 1M (cheapest route) |
| Market output pricing | $0.31 per 1M (cheapest route) |
| Reasoning tokens | Billed as output at $0.31 per 1M |
| Input modalities | text |
| Output modalities | text |
| Reasoning control | Yes, effort levels exposed |
| Tool support | Functions, tool choice, structured outputs |
| Tokenizer | DeepSeek |
| Provider routes | 6 endpoints |
| Artificial Analysis index | Not published |
| Speed | Not published |
| Knowledge cutoff | Not publicly stated |
Input and output support
| Modality | Input | Output | Billed at |
|---|---|---|---|
| Text | ✓ yes | ✓ yes | $0.209 in |
| Image | no | no | not supported |
| Audio | no | no | not supported |
| Video | no | no | not supported |
| File (PDF) | no | no | not supported |
02 · Overview
What is DeepSeek V3.2?
DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance.
It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism...
DeepSeek V3.2 carries a 164K token window, uses the DeepSeek tokenizer and is served by DeepSeek. On Hoodgen it sits in the Reasoning tier of the model menu next to its siblings, so you can switch mid conversation without losing the thread.
03 · Release
Release and what changed from the previous version
Released1 December 2025
| Property | DeepSeek V3.2 Exp | DeepSeek V3.2 | Change |
|---|---|---|---|
| Released | 2025 | 2025 | Same release window |
| Context | 164K | 164K | Unchanged |
| Max output | 147K | 147K | Unchanged |
| Input price | $0.27 | $0.209 | Cheaper |
| Output price | $0.41 | $0.31 | Cheaper |
| Input modalities | text | text | Unchanged |
| AA index | n/a | n/a | Newer benchmark run |
04 · Capabilities
Features and capabilities
◈
Thinking control
Reasoning effort can be dialled per request, so one model covers deep analysis and fast chat.
⚙
Function calling
Tools, tool choice and parallel calls, which is what agent loops need to act on results.
{ }
Structured outputs
Schema constrained JSON with response_format, so no parse and retry loop in your code.
⚡
Prompt caching
Cached input costs $0.13 per million instead of $0.209, so stable prefixes get cheaper.
⌁
Seeds
A seed can be pinned so the same prompt is reproducible across test runs.
05 · Reasoning
Thinking and reasoning capabilities
DeepSeek V3.2 can think before answering and exposes the effort level per request. Reasoning tokens bill as output at $0.31 per million, and the trace can be returned with include_reasoning.
| Level | First token | Cost per task | Best for |
|---|---|---|---|
| Measured | n/a | n/a | Artificial Analysis measures the default reasoning configuration |
| Low effort | Faster | Lower | Chat, autocomplete, classification |
| High effort | Slower | Higher | Debugging, planning, multi step agents |
07 · Use cases
What can you do with DeepSeek V3.2?
Debugging and root cause
Multi file failures where the model has to hold a causal chain before answering.
Architecture and planning
Design docs, migration plans and trade off analysis over long context.
Data analysis
Spreadsheets and logs read in one pass, then summarised with the numbers quoted back.
Agentic coding
Repo wide changes that need plan, edit and verify loops.
08 · Cost levers
Prompt caching and how to cut the bill
Cached input bills at $0.13 per million instead of $0.209. Keep the system prompt and documents first and the changing question last, so every call hits the cache.
09 · Pricing
DeepSeek V3.2 pricing
Official pricing
List price
$0.209 / 1M in
$0.31 / 1M out
- $0.13 cached input
- 164K context
- text input
Cheapest route
$0.209 / 1M in
$0.31 / 1M out
- 6 routes on OpenRouter
- What Hoodgen routes to first
- Same model, lower bill
Market pricing, every route
| Provider | Tier | Input | Output | Cache read | Uptime 24h |
|---|---|---|---|---|---|
| GMICloud | standard | $0.209 | $0.31 | $0.022 | 100.0% |
| SiliconFlow | standard | $0.259 | $0.8 | $0.14 | 100.0% |
| DeepInfra | standard | $0.26 | $0.38 | $0.13 | 100.0% |
| AtlasCloud | standard | $0.26 | $0.38 | $0.13 | 100.0% |
| Baidu | standard | $0.28 | $0.42 | $0.028 | n/a |
| DigitalOcean | standard | $0.3 | $0.96 | $0.09 | 100.0% |
API pricing and cost calculator
| Rate | Per call | Total | Credits |
|---|---|---|---|
| Cheapest route | $0.0024 | $1.20 | 1138 |
| List price | $0.0037 | $1.85 | 1756 |
10 · Developers
API and developer information
| Model ID | deepseek/deepseek-v3.2 |
| Endpoint | https://openrouter.ai/api/v1/chat/completions |
| Parameters | frequency_penalty · include_reasoning · logit_bias · logprobs · max_tokens · min_p · presence_penalty · reasoning · repetition_penalty · response_format · seed · stop · structured_outputs · temperature · tool_choice · tools · top_k · top_logprobs · top_p |
| Streaming | Yes, server sent events |
| Modality string | text->text |
| Max prompt tokens | 163,840 |
| Max completion tokens | 147,456 |
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek/deepseek-v3.2",
"messages": [{ "role": "user", "content": "Summarise this document." }],
"max_tokens": 1024
}'11 · Limitations
Model limitations
| Text only output | It cannot generate images, audio or video. Pair it with an image model on Hoodgen when you need visuals. |
| Output cap | 147,456 tokens per request. Long deliverables have to be chunked. |
| Benchmarks not published | Artificial Analysis does not publish an Intelligence Index for this release, so no benchmark score is claimed here. |
| Knowledge cutoff | The provider does not state a fixed cutoff for every release. Verify very recent events against search. |
12 · Comparison
DeepSeek V3.2 vs other DeepSeek models
| Model | Context | Input /1M | Output /1M | AA index | Released |
|---|---|---|---|---|---|
| DeepSeek V3.2 | 164K | $0.209 | $0.31 | n/a | 2025 |
| DeepSeek V4 Pro | 1.05M | $0.9 | $1.5 | 23 | 2026 |
| DeepSeek V3.1 Terminus | 164K | $0.27 | $1 | n/a | 2025 |
| DeepSeek V3.2 Exp | 164K | $0.27 | $0.41 | n/a | 2025 |
| DeepSeek V3 0324 | 164K | $0.25 | $1 | n/a | 2025 |
| DeepSeek V4.1 Flash 1017 | 1.05M | $0.038 | $0.5 | 33 | Oct 2026 |
13 · Privacy
Privacy and data policy
- On Hoodgen, prompts and attachments answer that request and stay in your conversation history so you can reopen it. They are not used to train models.
- Requests route through OpenRouter to DeepSeek and are covered by that provider's API terms.
- Do not paste secrets or personal data you would not send to any hosted model.
14 · FAQ
DeepSeek V3.2 questions, answered
DeepSeek V3.2 is a reasoning model from DeepSeek, released 1 December 2025. DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance.
✦
Start chatting with DeepSeek V3.2
Free to try on Hoodgen AI, with a live credit estimate before every send. No API key, no setup.