What is DeepSeek-R1-0528?
DeepSeek-R1-0528 is an updated version of DeepSeek's R1 reasoning model, released on May 28, 2025. It represents a meaningful upgrade over the original January 2025 R1, using more compute during inference and improved post-training techniques to push reasoning performance closer to top frontier models. It maintains the same MIT open-source license as the original R1, making it free to download and self-host.
What's Improved in R1-0528?
The key improvements are in complex reasoning accuracy and context length. R1-0528 supports a 164,000 token context window — more than double the original 64,000 — making it practical for longer documents and extended reasoning chains. Its post-training improvements yield stronger performance on AIME 2025, MATH-500, and competitive coding benchmarks, pushing it closer to OpenAI's o3 and Gemini 2.5 Pro on the hardest reasoning tasks.
DeepSeek describes R1-0528 as "tapping more compute and smarter post-training tricks, pushing its reasoning and inference to the brink of flagship models." In practice, users report noticeably sharper mathematical derivations and fewer reasoning errors on multi-step problems compared to the original R1.
DeepSeek-R1-0528 Pricing
The model costs $0.50 per million input tokens and $2.15 per million output tokens via DeepSeek's API. Cache hit pricing is $0.05 per million (90% discount). Reasoning tokens generated during chain-of-thought are counted as output tokens. For comparison, OpenAI's o1 charges approximately $15/$60 per million tokens — R1-0528 delivers comparable reasoning at roughly 1/28th the cost.
R1-0528 vs Original DeepSeek-R1
R1-0528 is strictly better than the original R1 in every measurable way: longer context, stronger reasoning, and comparable pricing. If you're using DeepSeek's reasoning model, R1-0528 is the version to use. The original R1 remains available but is no longer recommended for new deployments.
Frequently Asked Questions
Is DeepSeek-R1-0528 open source?
Yes. Like the original R1, the 0528 update is released under the MIT license. You can download the weights from Hugging Face and run inference on your own hardware without any usage restrictions.
How does R1-0528 compare to OpenAI's o3?
On most academic reasoning benchmarks, R1-0528 approaches o3 performance at a fraction of the API cost. On the hardest tasks (frontier math, advanced scientific reasoning), o3 still maintains an edge. For the vast majority of reasoning applications, R1-0528 is the most cost-effective choice available.