Anonymous Artificial Intelligence Model Ignites Rumours of DeepSeek's Next Major Release

2026-03-18 digital

Beijing, Wednesday 18 March 2026
A powerful, anonymous artificial intelligence model has surfaced online, boasting advanced reasoning capabilities and igniting global speculation that DeepSeek is quietly testing its next-generation architecture.

Cost Efficiency and Software Scalability

The economic appeal of DeepSeek’s architecture lies in its severe reduction of computational overhead, fundamentally altering software scalability. The company’s V3 model, which possesses 671 billion parameters, utilises a Mixture-of-Experts (MoE) architecture to activate only about 37 billion parameters per token [6]. This structural efficiency reduces active parameters by approximately 94.48% [6]. Consequently, DeepSeek’s Application Programming Interface (API) costs are remarkably low, priced at $0.28 per million tokens for input and $0.42 for output [5]. In stark contrast, ChatGPT’s API costs $2.50 for input and $15.00 for output [5], making ChatGPT’s output pricing 3471.429% more expensive. This aggressive pricing strategy allows Benelux-based startups and global enterprises to integrate advanced reasoning into legacy industries without prohibitive expenditure [GPT].

The Detection Dilemma and Competitive Dynamics

As foundation models become increasingly sophisticated, the cybersecurity and academic sectors face mounting challenges in distinguishing AI-generated content from human writing [3]. DeepSeek version 3.2 boasts formidable benchmark scores, including 88.50% on General Knowledge (MMLU) and 82.60% on Coding & Logic (HumanEval) [4]. A recent stress test utilising 72 pieces of high-complexity, long-form content generated by DeepSeek v3.2 highlighted critical vulnerabilities in legacy AI detection tools [4]. ZeroGPT, a widely used detection platform, correctly identified only 41 of the 72 samples, yielding an accuracy rate of 56.94% [4]. In contrast, the AI or Not platform correctly flagged 67 samples, achieving a 93.06% accuracy rate [4]. This performance gap of 36 points [4] underscores the rapid obsolescence of older detection infrastructures and the pressing need for continuous security upgrades [3][4].

Sources & Ecosystem Partners

Artificial intelligence Foundation models