DeepSeek R1 Surpasses GPT-4o on Frontier Benchmarks, Free API Launches
Overview
DeepSeek, the Shanghai-based AI lab, released DeepSeek R1, its largest reasoning model to date, in late June 2026. R1 achieves parity with OpenAI's GPT-4o and Claude 3.5 Sonnet on coding and math benchmarks while costing 99% less at inference. The announcement sent shockwaves through the AI industry and prompted OpenAI to slash GPT-4o pricing by 50% the following week.
DeepSeek R1 uses a Mixture-of-Experts (MoE) architecture with 671B total parameters but only 37B activated per token. Trained on a custom dataset of 10 trillion tokens, the model achieves a 92.3 pass@1 on HumanEval and 89.1% on GSM8K. The open-weight release on Hugging Face has been downloaded over 2 million times in two weeks.
The company also launched a free API for R1, offering 1 million tokens per month at no cost, with paid tiers starting at $0.14 per million input tokens. This pricing has made R1 the default backbone for many open-source agent frameworks.
Key Features
- 671B MoE architecture: 37B activated per token, low inference cost.
- Open weights: Released under DeepSeek License, allows commercial use.
- 1M context: Supports long-context reasoning tasks.
- Tool-use support: Native function calling with JSON schema.
- Free API: 1M tokens/month free tier.
Installation / Access
# Via DeepSeek API
pip install deepseek
from deepseek import DeepSeekClient
client = DeepSeekClient(api_key="your-key")
response = client.chat.completions.create(
model="deepseek-r1",
messages=[{"role": "user", "content": "Explain the Transformer architecture."}],
max_tokens=2048
)
print(response.choices[0].message.content)
Benchmarks
| Benchmark | DeepSeek R1 | GPT-4o | Claude 3.5 Sonnet |
|---|---|---|---|
| HumanEval | 92.3 | 91.0 | 89.6 |
| GSM8K | 89.1 | 87.6 | 88.2 |
| MMLU | 87.4 | 88.7 | 88.2 |
| Inference Cost | $0.14/1M | $10.00/1M | $3.00/1M |
Pros
- Lowest cost among frontier models
- Open weights for self-hosting
- Strong reasoning on math and code
- Free API tier for experimentation
Cons
- English-only training data for now
- Chinese bias in some outputs
- Less creative writing quality than Claude
- No official vision support yet