How much does DeepSeek V3.2 cost?

DeepSeek V3.2 costs $0.2088/M input tokens and $0.3096/M output tokens, with separate rates for Cache Read at $0.0216/M tokens.

What is the context length of DeepSeek V3.2?

DeepSeek V3.2 has a 163,840 token context window.

Does DeepSeek V3.2 support tool calling and structured outputs?

Yes. DeepSeek V3.2 accepts tools and tool_choice for function calling on 4 of the 6 providers serving it, and requests that send tools are routed to those providers. It also supports structured outputs via a JSON schema in response_format.

Which providers serve DeepSeek V3.2?

DeepSeek V3.2 is served by 6 providers on OpenRouter: GMICloud, SiliconFlow, DeepInfra, AtlasCloud, Baidu Qianfan, and DigitalOcean. Requests are routed to the best available provider, with automatic failover to the others, and you can pin or exclude providers with provider routing.

When was DeepSeek V3.2 released?

DeepSeek V3.2 was released on December 1, 2025.

DeepSeek: DeepSeek V3.2

deepseek/deepseek-v3.2

Model weights

Compare

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 that reduces training and inference cost while preserving quality in long-context scenarios. A scalable reinforcement learning post-training framework further improves reasoning, with reported performance in the GPT-5 class, and the model has demonstrated gold-medal results on the 2025 IMO and IOI. V3.2 also uses a large-scale agentic task synthesis pipeline to better integrate reasoning into tool-use settings, boosting compliance and generalization in interactive environments.

Users can control the reasoning behaviour with the reasoning enabled boolean. Learn more in our docs

Modalities

In / out price

28% off

$0.2088 / $0.3096per 1M

Context

164K

Released

Dec 1, 2025

Providers

Different companies host the same model. OpenRouter routes your request to one of them based on the routing mode you pick: Balanced (price + speed), Nitro (fastest), or Floor (cheapest).

Latency / throughput


28% off	$0.29$0.2088	$0.43$0.3096	$0.03$0.0216	1.23s	53 tok/s	100.00%
	$0.28	$0.42	$0.028	1.00s	35 tok/s	99.91%
	$0.26	$0.38	$0.13	0.59s	34 tok/s	99.64%
	$0.30	$0.96	$0.09	2.80s	27 tok/s	97.89%
	$0.259	$0.80	$0.14	2.03s	19 tok/s	99.91%
	$0.26	$0.38	$0.13	2.05s	26 tok/s	100.00%

Input $0.2088Output $0.3096
28% off
53 tok/s100.00% uptime
Input $0.28Output $0.42
35 tok/s99.91% uptime
Input $0.26Output $0.38
34 tok/s99.64% uptime
Input $0.30Output $0.96
27 tok/s97.89% uptime
Input $0.259Output $0.80
19 tok/s99.91% uptime
Input $0.26Output $0.38
26 tok/s100.00% uptime

Pricing

The average price customers actually pay for this model, next to the prices providers post. Caching and discounts mean the price actually paid is often well below the listed one.

Weighted average

Input

$0.1381/M tokens

Output

$0.3905/M tokens

DeepSeek V3.2 — Price history

Price history

No effective price data for this selection

$0.0871	$0.309	65.0%	15.0%
$0.1094	$0.419	67.7%	41.0%
$0.1734	$0.379	66.6%	28.5%
$0.1760	$0.799	69.7%	5.3%
$0.1946	$0.959	50.2%	2.2%
$0.2048	$0.379	42.4%	8.1%

Effective in $0.0871Effective out $0.309
Token share
15.0%
Effective in $0.1094Effective out $0.419
Token share
41.0%
Effective in $0.1734Effective out $0.379
Token share
28.5%
Effective in $0.1760Effective out $0.799
Token share
5.3%
Effective in $0.1946Effective out $0.959
Token share
2.2%
Effective in $0.2048Effective out $0.379
Token share
8.1%

Performance

Throughput is how fast the model writes (tokens per second — higher is better). Latency is total round-trip time (lower is better). TTFT is time-to-first-token — how long before you see anything appear (lower is better).

Throughput

53tok/s

P50, best across providers

Latency

0.59s

P50, best provider

Latency / throughput

AutoExacto benchmarks

GPQA DiamondTAU-BenchDigitalOcean82.5%--Baidu Qianfan81.9%--GMICloud88.1%69.5%Venice82.2%74.2%Friendli82.1%74.0%

+8 more providers

Uptime

Uptime is the percentage of the past 3 days that at least one provider was responding to requests. Availability is the percentage of time that inference was successfully served. OpenRouter continuously monitors and uses the next-best provider when one returns an error.

Uptime (3d)

100.00%

Availability (3d)

99.45%

Availability over the last 3 days

Last 72 hours

Availability 99.45%

3 Days Ago2 Days AgoYesterdayNow

Availability over the last 24 hours

OpenRouter Availability

99.46%

Without Routing

97.23%

When an error occurs in an upstream provider, we can recover by routing to another healthy provider, if your request filters allow it. You can access per-provider uptime data programmatically through the Endpoints API. Learn more about our load balancing and provider routing options.

Benchmarks

Scores on standardized evaluations and this model’s own outputs on shared benchmark prompts. Higher percentages are better — and each rank shows where this model lands among the models evaluated on that source.

Benchmark score summary for DeepSeek: DeepSeek V3.2 (Artificial Analysis, Design Arena and Vals AI)
Source	Benchmark	Score
Artificial Analysis	DeepSeek V3.2 (Reasoning) Coding Index	44.2
Artificial Analysis	DeepSeek V3.2 (Reasoning) GPQA Diamond	84.0%
Artificial Analysis	DeepSeek V3.2 (Reasoning) HLE	24.6%
Artificial Analysis	DeepSeek V3.2 (Reasoning) IFBench	60.7%
Artificial Analysis	DeepSeek V3.2 (Reasoning) τ²-Bench Telecom	90.6%
Artificial Analysis	DeepSeek V3.2 (Reasoning) AA-LCR	73.3%
Artificial Analysis	DeepSeek V3.2 (Reasoning) GDPval-AA	10.0%
Artificial Analysis	DeepSeek V3.2 (Reasoning) CritPt	2.9%
Artificial Analysis	DeepSeek V3.2 (Reasoning) LiveCodeBench	86.2%
Artificial Analysis	DeepSeek V3.2 (Reasoning) Terminal-Bench Hard	35.6%
Artificial Analysis	DeepSeek V3.2 (Reasoning) Terminal-Bench 2.1	46.8%
Artificial Analysis	DeepSeek V3.2 (Reasoning) MMLU-Pro	86.2%
Artificial Analysis	DeepSeek V3.2 (Reasoning) AIME 2025	92.0%
Design Arena	DeepSeek-V3.2 Models Arena 3D Elo	1112
Design Arena	DeepSeek-V3.2 Models Arena Asciiart Elo	1077
Design Arena	DeepSeek-V3.2 Models Arena Code Categories Elo	1153
Design Arena	DeepSeek-V3.2 Models Arena Data Visualization Elo	1155
Design Arena	DeepSeek-V3.2 Models Arena Game Development Elo	1121
Design Arena	DeepSeek-V3.2 Models Arena SVG Elo	1022
Design Arena	DeepSeek-V3.2 Models Arena UI Component Elo	1139
Design Arena	DeepSeek-V3.2 Models Arena Website Elo	1167
Vals AI	DeepSeek V3.2 (Thinking) AIME	84.6%
Vals AI	DeepSeek V3.2 (Thinking) CaseLaw v2	55.4%
Vals AI	DeepSeek V3.2 (Thinking) CorpFin v2	51.0%
Vals AI	DeepSeek V3.2 (Thinking) Finance Agent v1.1	2.3%
Vals AI	DeepSeek V3.2 (Thinking) GPQA Diamond	80.3%
Vals AI	DeepSeek V3.2 (Thinking) IOI v1	10.7%
Vals AI	DeepSeek V3.2 (Thinking) LiveCodeBench	80.7%
Vals AI	DeepSeek V3.2 (Thinking) LegalBench	76.1%
Vals AI	DeepSeek V3.2 (Thinking) MedQA	93.9%
Vals AI	DeepSeek V3.2 (Thinking) MGSM	86.0%
Vals AI	DeepSeek V3.2 (Thinking) MMLU Pro	84.9%
Vals AI	DeepSeek V3.2 (Thinking) ProofBench v1	4.0%
Vals AI	DeepSeek V3.2 (Thinking) SWE-bench	67.6%
Vals AI	DeepSeek V3.2 (Thinking) TaxEval v2	68.2%
Vals AI	DeepSeek V3.2 (Thinking) Terminal-Bench 2.0	36.0%
Vals AI	DeepSeek V3.2 (Thinking) Vals Index	26.3%
Vals AI	DeepSeek V3.2 (Thinking) Vibe Code Bench v1.1	5.1%

Apps

Public apps that send the most traffic to this model. Good signal for what real production workloads look like — and a hint at which use cases this model is best suited for.

Chatbox AI

On-device AI copilot

152Btokens

Janitor AI

Janitor AI is a chatbot platform where users create and chat with custom AI characters for interactive roleplay, storytelling, and immersive fiction.

145Btokens

Hermes Agent

Hermes Agent is an open-source, self-improving AI agent by Nous Research that runs persistently with memory across sessions, and builds reusable skills from experience. It comes with 40+ built-in tools, including web search, browser automation, and vision, plus scheduled automations and subagents.

17.4Btokens

Sophia's LoreBary

Organize and enhance your roleplay creations.

16.4Btokens

SillyTavern

SillyTavern is the LLM frontend for power users, a chat interface that connects to any model and gives deep control through character creation, roleplay, and prompt customization.

13.4Btokens

Rest of OpenRouter

8.65B

Top 5 Apps

3.55B

Not currently available through EU in-region routing. In-region routing docs

More models from DeepSeek

DeepSeek V4.1 Flash 1017

DeepSeek V4.1 Flash 1017 is the October 17, 2026 snapshot of DeepSeek V4.1 Flash, served by DeepSeek under the deepseek-flash API model name. V4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on output from a 552B-parameter backbone, an asymmetric split that keeps per-token compute low relative to the model's total size. Image understanding is native to the architecture, with visual and text embeddings trained jointly from the start of pre-training rather than added afterward as in the earlier experimental V4 Flash Vision Exp.

It is suited for coding, terminal, and computer-use agents, along with long-horizon tasks that must run to completion across many steps and long-context analysis. Compressed KV caching cuts cache memory to roughly a quarter of the previous Flash generation, significantly reducing costs on agentic workloads. DeepSeek positions it as the cost-efficient tier of the V4.1 family and reports that it exceeds V4 Pro on performance, speed, and task completion time.

Text1.0M context$0.038 / $0.50

DeepSeek V4.1 Flash 1017 (batch)

Text1.0M context$0.05 / $0.15

DeepSeek Pro Latest

This model always redirects to the latest model in the DeepSeek Pro family.

Text1.0M context

DeepSeek Flash Latest

This model always redirects to the latest model in the DeepSeek Flash family.

Text1.0M context

DeepSeek V4.1 Flash

DeepSeek V4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on output from a 552B-parameter backbone, an asymmetric split that keeps per-token compute low relative to the model's total size. Image understanding is native to the architecture, with visual and text embeddings trained jointly from the start of pre-training rather than added afterward as in the earlier experimental V4 Flash Vision Exp.

Text1.0M context$0.0507 / $0.2028

DeepSeek V4.1 Flash (batch)

Text1.0M context$0.112 / $0.336

DeepSeek V4 Flash Vision Exp

DeepSeek V4 Flash Vision Exp is an experimental vision-enabled version of DeepSeek V4 Flash 0731 from DeepSeek, adding image understanding while matching the base model on text capabilities including agents, reasoning, and world knowledge. It is a sparse mixture-of-experts model with 13B active parameters out of 284B total.

It is suited for document and chart understanding, visual question answering, and multimodal agent workflows that interleave text and images.

Text1.0M context$0.2156 / $0.6468

DeepSeek V4 Pro 0813

DeepSeek V4 Pro 0813 is a large-scale mixture-of-experts model from DeepSeek. This is the GA release of DeepSeek V4 Pro.

Text1.0M context$0.0999 / $8

DeepSeek V4 Flash Latest

This model always redirects to the latest model in the DeepSeek V4 Flash family.

Text1.0M context

DeepSeek V4 Flash 0731

DeepSeek V4 Flash 0731 is a sparse mixture-of-experts model from DeepSeek, with 13B active parameters out of 284B total. This re-post-trained revision is suited for coding, reasoning, and agent workflows. This is the GA release of DeepSeek V4 Flash.

Text1.0M context$0.0434 / $0.30

DeepSeek V4 Flash 0731 (batch)

Text1.0M context$0.048 / $0.144

DeepSeek V4 Pro 0423

DeepSeek V4 Pro is a large-scale Mixture-of-Experts model from DeepSeek with 1.6T total parameters and 49B activated parameters, supporting a 1M-token context window. It is designed for advanced reasoning, coding, and long-horizon agent workflows, with strong performance across knowledge, math, and software engineering benchmarks.

Built on the same architecture as DeepSeek V4 Flash, it introduces a hybrid attention system for efficient long-context processing. Reasoning efforts high and xhigh are supported; xhigh maps to max reasoning. It is well suited for complex workloads such as full-codebase analysis, multi-step automation, and large-scale information synthesis, where both capability and efficiency are critical

Text1.0M context$0.3327 / $0.6654

DeepSeek V4 Flash 0423

DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts model from DeepSeek with 284B total parameters and 13B activated parameters, supporting a 1M-token context window. It is designed for fast inference and high-throughput workloads, while maintaining strong reasoning and coding performance.

The model includes hybrid attention for efficient long-context processing. Reasoning efforts high and xhigh are supported; xhigh maps to max reasoning. It is well suited for applications such as coding assistants, chat systems, and agent workflows where responsiveness and cost efficiency are important.

Text1.0M context$0.02226 / $0.04452

DeepSeek V3.2 Speciale

DeepSeek-V3.2-Speciale is a high-compute variant of DeepSeek-V3.2 optimized for maximum reasoning and agentic performance. It builds on DeepSeek Sparse Attention (DSA) for efficient long-context processing, then scales post-training reinforcement learning to push capability beyond the base model. Reported evaluations place Speciale ahead of GPT-5 on difficult reasoning workloads, with proficiency comparable to Gemini-3.0-Pro, while retaining strong coding and tool-use reliability. Like V3.2, it benefits from a large-scale agentic task synthesis pipeline that improves compliance and generalization in interactive environments.

Text131K context

DeepSeek V3.2

Users can control the reasoning behaviour with the reasoning enabled boolean. Learn more in our docs

Text164K context$0.2088 / $0.3096

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 designed to improve training and inference efficiency in long-context scenarios while maintaining output quality. Users can control the reasoning behaviour with the reasoning enabled boolean. Learn more in our docs

The model was trained under conditions aligned with V3.1-Terminus to enable direct comparison. Benchmarking shows performance roughly on par with V3.1 across reasoning, coding, and agentic tool-use tasks, with minor tradeoffs and gains depending on the domain. This release focuses on validating architectural optimizations for extended context lengths rather than advancing raw task accuracy, making it primarily a research-oriented model for exploring efficient transformer designs.

Text164K context$0.27 / $0.41

DeepSeek V3.1 Terminus

DeepSeek-V3.1 Terminus is an update to DeepSeek V3.1 that maintains the model's original capabilities while addressing issues reported by users, including language consistency and agent capabilities, further optimizing the model's performance in coding and search agents. It is a large hybrid reasoning model (671B parameters, 37B active) that supports both thinking and non-thinking modes. It extends the DeepSeek-V3 base with a two-phase long-context training process, reaching up to 128K tokens, and uses FP8 microscaling for efficient inference. Users can control the reasoning behaviour with the reasoning enabled boolean. Learn more in our docs

The model improves tool use, code generation, and reasoning efficiency, achieving performance comparable to DeepSeek-R1 on difficult benchmarks while responding more quickly. It supports structured tool calling, code agents, and search agents, making it suitable for research, coding, and agentic workflows.

Text164K context$0.27 / $1

DeepSeek V3.1

DeepSeek-V3.1 is a large hybrid reasoning model (671B parameters, 37B active) that supports both thinking and non-thinking modes via prompt templates. It extends the DeepSeek-V3 base with a two-phase long-context training process, reaching up to 128K tokens, and uses FP8 microscaling for efficient inference. Users can control the reasoning behaviour with the reasoning enabled boolean. Learn more in our docs

It succeeds the DeepSeek V3-0324 model and performs well on a variety of tasks.

Text164K context$0.25 / $0.95

DeepSeek V3.1 Base

This is a base model, trained only for raw next-token prediction. Unlike instruct/chat models, it has not been fine-tuned to follow user instructions. Prompts need to be written more like training text or examples rather than simple requests (e.g., “Translate the following sentence…” instead of just “Translate this”).

DeepSeek-V3.1 Base is a 671B parameter open Mixture-of-Experts (MoE) language model with 37B active parameters per forward pass and a context length of 128K tokens. Trained on 14.8T tokens using FP8 mixed precision, it achieves high training efficiency and stability, with strong performance across language, reasoning, math, and coding tasks.

Text164K context

R1 Distill Qwen 7B

DeepSeek-R1-Distill-Qwen-7B is a 7 billion parameter dense language model distilled from DeepSeek-R1, leveraging reinforcement learning-enhanced reasoning data generated by DeepSeek's larger models. The distillation process transfers advanced reasoning, math, and code capabilities into a smaller, more efficient model architecture based on Qwen2.5-Math-7B. This model demonstrates strong performance across mathematical benchmarks (92.8% pass@1 on MATH-500), coding tasks (Codeforces rating 1189), and general reasoning (49.1% pass@1 on GPQA Diamond), achieving competitive accuracy relative to larger models while maintaining smaller inference costs.

Text131K context

DeepSeek: DeepSeek V3.2

Providers