According to Artificial Analysis, the model maintains a size of 284 billion total parameters but uses a mixture-of-experts architecture where only 13 billion parameters are active during inference—the process of the AI generating a response.
The update significantly shifts the competitive landscape for cost-efficient AI infrastructure by maintaining the same pricing as the previous version while delivering higher performance.
Artificial Analysis reports that DeepSeek’s model is approximately 60% cheaper per task than GPT-5.6 Luna, even after recent price cuts by OpenAI.
This cost advantage is largely driven by a 98% discount on "cache hits," a mechanism where the system identifies and reuses previously processed data to save on computing power.
These savings, combined with a 12% reduction in total output tokens used compared to its predecessor, place the model on the industry "Pareto frontier," representing the best available balance between intelligence and cost.
Technical evaluations show that the performance jump is primarily due to improved reliability rather than a broader knowledge base.
The model’s higher score on the Omniscience Index was driven entirely by a reduction in "hallucinations"—instances where an AI generates false information—as its raw accuracy remained unchanged.
The model also saw a substantial rise in its rating for "agentic" tasks, which involve the AI performing complex, real-world work autonomously.
While currently available through DeepSeek’s API, the company is expected to release the model’s full weights in the coming weeks, allowing developers to run the software on their own hardware.