DeepSeek v3.2
DeepSeek v3.2 via the Vercel AI Gateway: high efficiency with superior reasoning and agent performance. 675B MoE. Served only by zero-data-retention providers. Does not support web search or image input.
You always get the exact model you pick — we never silently route you to another.
About DeepSeek v3.2
DeepSeek V3.2 is built for one thing above most others: math. With a 93.1% score on AIME 2025 and strong results on GPQA Diamond (82.4%), it outperforms several closed-source models on mathematical olympiad-style problems and scientific reasoning tasks — at a fraction of the price, coming in around $0.28 per million input tokens through the official API. Users consistently praise the cost-to-capability ratio and appreciate its open-source availability, which sets it apart from the proprietary alternatives it competes with on benchmarks. The extended 128K–163.8K token context window makes it genuinely useful for long-document analysis and multi-step technical problems. That said, the experience is not uniformly polished: inference is slower than competing models, and users note that real-world performance in agent and tool-use scenarios trails what the benchmark numbers suggest. Complex multi-step autonomous workflows are where it struggles most. If your work centers on quantitative reasoning, cost-sensitive API usage, or open research — and you can tolerate the slower pace — V3.2 is a well-priced, capable choice. Note that deprecation is scheduled for July 24, 2026.
Best for
- Mathematical and scientific problem-solving, including advanced math competitions and chemistry or physics reasoning tasks
- High-volume, cost-sensitive API workloads where keeping token costs low matters more than raw speed
- Long-document analysis and summarization, leveraging its 128K–163.8K token context window
- Code review and medium-complexity technical writing backed by solid benchmark performance on LiveCodeBench
- Open-source research and development where model transparency and accessibility are priorities
Specifications
| Provider | DeepSeek |
|---|---|
| Released | 2025-12 |
| Intelligence | Medium |
| Context window | 128,000 tokens |
| Max output | 64,000 tokens |
| Knowledge cutoff | Not specified in available sources |
| Input price | $0.62 / 1M tokens |
| Output price | $1.85 / 1M tokens |
| Request cost | 3 base requests |
| Plan tier | Base |
| Features | Function calling supported, Served through the Vercel AI Gateway (deepseek/deepseek-v3.2), Zero data retention and no prompt training |
| Model ID | deepseek-v3p2 |
Frequently asked questions
Through the official DeepSeek API it costs $0.28 per million input tokens and $0.42 per million output tokens. Blended across providers, Artificial Analysis puts the average at around $0.54 per million tokens.
The standard context window is 163.8K tokens (128K on some third-party providers), with a maximum output of up to 65.5K tokens depending on the provider.
Mathematical and scientific reasoning. It scored 93.1% on AIME 2025 and 82.4% on GPQA Diamond, making it one of the stronger models for quantitative tasks at its price point.
Agent reliability and multi-step tool use. Users report that autonomous CLI and SWE agent performance is inconsistent, and complex tool orchestration workflows lag behind GPT-5 and Gemini. Inference speed is also notably slower than competing models.
V3.2-Exp is an earlier experimental reasoning variant with slightly different performance characteristics. V3.2 is the official successor and the stable release.
Yes, but it is being phased out. Deprecation is scheduled for July 24, 2026, as DeepSeek transitions to V4 models.