The global artificial intelligence landscape is undergoing a significant reassessment of pricing policies, leading to changes in infrastructure choices among developers. Models characterized by high economic efficiency, such as DeepSeek, Zhipu AI, and Qwen, have accounted for over 20% of the weekly token volume on the OpenRouter platform.
According to OpenRouter, an aggregator of numerous AI models for developers, DeepSeek, Zhipu AI, and Qwen models have consistently exceeded 20% of weekly token consumption since February 2026, reaching peaks of 30%, which is significantly higher than the average of 7% over the past year.
The cost-to-performance ratio is fundamentally changing developers' decisions. For instance, the input token cost of DeepSeek V4 is approximately 5% lower than that of GPT-5.5, while output costs are around 7%. The Zhipu GLM-5.2 model has an input price approximately 28% lower than GPT-5.5, with output expenses at about 15%. When compared to Anthropic Sonnet 5, GLM-5.2 is 30% cheaper for input data and 56% cheaper for output.
These price differences provide substantial annual savings when used in high-volume applications. Benchmarks also confirm the competitiveness of these models. DeepSeek V4 ProMax demonstrated comparable results on LiveCodeBench, Terminal Bench, and BrowseComp. Furthermore, the Zhipu GLM-5.1 Thinking model surpassed Claude Opus 4.6 and GPT-5.4 in the SWE Pro test, while GLM-5.2 trails Claude Opus 4.8 by only 1% in FrontierSWE, yet costs about five times less.
In addition to the cost advantage, open models from DeepSeek, Zhipu AI, and Qwen offer enterprises flexibility through self-deployment capabilities and data sovereignty. These capabilities are unavailable in closed models, aligning with the growing demand from companies for controlled AI architectures in public, private, and hybrid environments.
The founder of an American AI company shared on the social network X that switching from Anthropic to DeepSeek V4 allowed them to reduce costs by millions of dollars while maintaining performance across many use cases. This cost-driven migration trend is gaining popularity in the global developer ecosystem, signaling a shift in market focus from 'who is the most powerful' to 'who provides the most economically efficient intelligence.'
