deepseek-reasoner
by deepseek
Pricing
Input
$0.04 / 1M tokens
Output
$0.20 / 1M tokens
DeepSeek-Reasoner Release by DeepSeek AI
Overview
DeepSeek-Reasoner is a reasoning-first language model released by DeepSeek AI in mid-2024. It explicitly separates chain-of-thought reasoning from final answers, enabling interpretable and auditable outputs.
- Release Date: May 28, 2024
- Chain-of-Thought: Auto-generated and accessible before final output
- Context Length: 64,000 tokens (CoT reasoning doesn’t count toward this)
- Final Output: Max 64,000 tokens including reasoning
Features
- Explicit reasoning pathway (CoT output)
- High coding, math, and logic strength
- Part of the DeepSeek-R1 family with advanced RL tuning
- OpenAI-compatible API
- Transparent, auditable output generation
Benchmarks
| Task | DeepSeek-Reasoner | Claude 3.5 Haiku | GPT-4o Mini |
|---|---|---|---|
| MMLU (reasoning) | 84.5% | 82.5% | 82.0% |
| GSM8K (math) | 92.4% | 89.0% | 87.0% |
| HumanEval (coding) | 89.1% | 87.5% | 87.2% |
| MMMU (multimodal) | N/A | 59.2% | 59.4% |
Pricing
| Type | Price per 1M tokens |
|---|---|
| Input | $0.10 |
| Output (includes CoT) | $0.50 |
Reasoning tokens (CoT) are priced equally with final answers.
Use Cases
- Transparent academic tutors with visible reasoning
- Compliance-focused AI tools with auditable decisions
- Step-by-step logic generation for math/code solvers
- Research assistants and scientific automation
- AI systems requiring explainability and trust
Safety and Stability
- Follows strict instruction-tuning methods
- Internal reasoning can be reviewed or filtered
- CoT generation improves reliability of answers
- No browsing; offline inference only
Limitations
- Slower response due to dual-stage output
- No multimodal (vision/audio) support
- Relatively new; limited third-party benchmarks
License
DeepSeek-Reasoner is commercially usable via API or local deployment under DeepSeek’s open license. Available via API or HuggingFace.