OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Input | |
| Output | |
| Benchmark | Score | What it measures | Source | Retrieved |
|---|---|---|---|---|
| Intelligence | 14.5 | Composite score across standardised reasoning, knowledge and problem-solving evaluations, measured independently by Artificial Analysis | Source | September 12, 2026 |
| Math | 96.7 | Composite score across standardised mathematics evaluations, measured independently by Artificial Analysis | Source | September 12, 2026 |
| AIME 2025 | 96.0% | Competition mathematics (AIME), 2025 | Source | September 20, 2026 |
| GPQA Diamond | 85.7 | Google-proof graduate science questions (hardest subset) | Source | September 20, 2026 |
| LiveCodeBench | 88.7% | Contamination-free competitive programming problems | Source | September 20, 2026 |
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
DeepSeek-V3.2 Speciale This page | ||||
| Reasoning + agents | ||||
| Balanced coding + agents | ||||
| Long-context, multimodal & agentic workflows | ||||
| Reasoning + agents |
Yes, DeepSeek-V3.2 Speciale can stream responses as they are generated.
Yes, DeepSeek-V3.2 Speciale supports both function calling and structured outputs.
DeepSeek-V3.2 Speciale was built by DeepSeek.
It supports function calling, streaming, and structured outputs, which are features commonly needed to build agentic workflows.
It is accessed via the v1 endpoint.
