# DeepSeek-R1

> DeepSeek's flagship reasoning model with chain-of-thought. Strong math, coding, and complex problem solving.

**ID**: `deepseek/deepseek-r1`  
**Creator**: DeepSeek  
**Category**: text  
**Context**: 64K tokens  
**Pricing**: $0.70 / $2.50 per 1M · billed by your provider (BYOK)  
**Released**: 2026-07-08  
**Web page**: https://anyrouter.dev/model/deepseek/deepseek-r1

**Input modalities**: text  
**Output modalities**: text  

**Capabilities**: chat, reasoning, streaming, function-calling

**Aliases**: `deepseek/DeepSeek-R1`, `deepseek-ai/DeepSeek-R1`

## Usage

**Endpoint**: `POST https://anyrouter.dev/api/v1/chat/completions`

```bash
curl https://anyrouter.dev/api/v1/chat/completions \
  -H "Authorization: Bearer $ANYROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "deepseek/deepseek-r1",
  "messages": [
    {
      "role": "user",
      "content": "Say hi in 3 words."
    }
  ]
}'
```

## Providers

| Provider | Input | Output |
| --- | --- | --- |
| siliconflow-byok | $0.70/M (list) | $2.50/M (list) |
| cline-byok | $0.70/M (list) | $2.50/M (list) |
| vercel-byok | $1.35/M (list) | $5.40/M (list) |
| openrouter-byok | $0.70/M (list) | $2.50/M (list) |
| huggingface-byok | $0.70/M (list) | $2.50/M (list) |

BYOK routes are billed by your own provider key — AnyRouter charges $0. Rates marked (list) are the provider's published list price.

## Benchmarks

- **GPQA Diamond**: 0.7
- **HLE**: 0.1
- **τ²-Bench Telecom**: 0.1
- **Aider Polyglot**: 0.7
- **SciCode**: 0.4
- **MATH Level 5**: 0.9
- **OTIS Mock AIME**: 0.5

Benchmark data from [Aider](https://aider.chat/docs/leaderboards/) (Apache-2.0) and [Epoch AI](https://epoch.ai/benchmarks) (CC-BY 4.0).

## Model card

- Architecture: mixture of experts [1]
- Total parameters: 671B [1,2]
- Active parameters: 37B [1,2]
- Native context: 131072 tokens [1]
- License: MIT, open weights (https://huggingface.co/deepseek-ai/DeepSeek-R1) [1]

### Strengths [1]

- Comparable to OpenAI o1 on math, code and reasoning
- Reasoning trained with large-scale reinforcement learning
- Distilled into dense models from 1.5B to 70B

### Known limits [2]

- Weak structured output; cannot call tools such as search
- Mixes languages outside Chinese and English
- Sensitive to prompts; few-shot prompting lowers quality

### Use cases [1]

- Math, code and logic reasoning
- Distilling reasoning into smaller models

### API notes [1]

- Model card advises no system prompt; put instructions in the user turn
- Model card recommends temperature 0.6 (range 0.5 to 0.7)

### Links

- model card: https://huggingface.co/deepseek-ai/DeepSeek-R1
- paper: https://arxiv.org/abs/2501.12948

### Sources

1. [deepseek-ai/DeepSeek-R1 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1/raw/main/README.md) (accessed 2026-10-14)
2. [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://arxiv.org/abs/2501.12948) (accessed 2026-10-14)

## Source

- Upstream docs: https://docs.siliconflow.com