Competition over the cost of artificial intelligence is intensifying as OpenAI and Anthropic lower the price of running powerful AI models, while Chinese developers such as DeepSeek and Moonshot AI continue to challenge Silicon Valley with significantly cheaper alternatives.

The battle is increasingly moving beyond benchmark scores. For companies processing millions or billions of tokens, the cost of deploying an AI model can matter almost as much as raw performance, giving lower-priced Chinese systems a growing advantage in high-volume applications.

OpenAI made one of the clearest moves on July 30, cutting the API price of GPT-5.6 Luna by 80 percent. The model now costs $0.20 per million input tokens and $1.20 per million output tokens, while the more capable GPT-5.6 Terra is priced at $2 for input and $12 for output.

OpenAI said improvements in model and infrastructure efficiency made the lower pricing possible. The company is positioning cheaper models for tasks such as document analysis, customer interaction classification and routine software work, where enormous volumes can quickly turn small differences in token prices into substantial costs.

Chinese models are forcing prices lower

Chinese AI developers have spent the past several years aggressively competing on the relationship between performance and price.

DeepSeek became one of the most prominent examples after demonstrating that capable AI systems could be offered at a fraction of the cost associated with many leading Western models. That price advantage remains significant in 2026.

Reuters reported that DeepSeek’s V4-Flash was priced at about $0.14 per million input tokens and $0.28 per million output tokens at the time of comparison.

The difference becomes even clearer when the cost of completing the same benchmark is considered. According to figures from Artificial Analysis cited by Reuters, one benchmark run cost roughly three cents with DeepSeek V4-Flash, compared with about $1.86 using OpenAI’s GPT-5.6 Sol and $3.15 with Anthropic’s Claude Fable 5.

Those figures do not mean the models provide identical performance. A cheaper system may require more tokens, additional prompting or multiple attempts to complete the same task. But the gap is large enough to influence purchasing decisions, particularly for companies operating AI products at scale.

Other Chinese developers are also strengthening the competitive pressure. Alibaba and Moonshot AI have introduced increasingly capable models while emphasizing lower prices and, in several cases, open-weight systems that developers can deploy and modify outside a single proprietary platform.

Anthropic is also moving down the price curve

Anthropic has likewise adjusted the economics of its model lineup.

Claude Opus 5 is currently priced at $5 per million input tokens and $25 per million output tokens, half the price of Claude Fable 5. Claude Sonnet 5 remains positioned below both at $2 for input and $10 for output.

The result is an AI market increasingly divided into pricing tiers rather than a simple contest over which company can build the most powerful model.

OpenAI and Anthropic can still charge premiums for their most advanced systems, but they also need cheaper models capable of handling routine enterprise workloads without allowing computing costs to grow uncontrollably.

For businesses, that distinction is becoming increasingly important. A customer-service platform, coding assistant or document-processing system may execute millions of requests every day. Moving repetitive tasks to a model that is slightly less capable but dramatically cheaper can produce substantial savings.

The pressure from Chinese competitors therefore reaches beyond headline model rankings. It is forcing leading US AI companies to compete on price-performance, a metric likely to become increasingly important as artificial intelligence shifts from experimental deployments to large-scale everyday infrastructure.

The next stage of the AI race may consequently depend less on which company produces the highest-scoring model and more on a simpler question: how much useful work can an AI system deliver for every dollar spent on computing?