“From Model Performance to Efficiency”: Cost Competition Heats Up as AI Capabilities Converge—Can Cheap Electricity Give China the Edge?

Narrowing Performance Gap Among Frontier AI Models
Growing Adoption of Task-Specific Model Combinations and Orchestration
Shift in Market Competition From Model Performance to Cost per Task

The rules of competition in the artificial intelligence (AI) market are changing. As the race to scale up parameter counts in massive models reaches its limits, enterprise customers are restructuring their AI adoption strategies around cost reduction, rapidly shifting competition among developers from performance to cost. This cost-centric competitive landscape, however, is likely to favor China, with its low-cost electricity and highly efficient models, potentially challenging US leadership in AI.

Optimal Task-Specific Model Combinations Taking Priority Over Top-Performing Models

In an interview with CNBC on July 12, local time, Perplexity CEO Aravind Srinivas said, “A model alone can no longer be a product,” adding that “the key is an orchestration system that connects the appropriate models and tools for each situation.” As relatively inexpensive models increasingly handle customer service and internal workflow automation, while high-performance models are reserved for complex coding and reasoning tasks, the competitive benchmark in the AI market is shifting from model performance to operational efficiency, he explained.

“The right approach to AI adoption is to use the model best suited to the task at hand,” Srinivas continued. Because each task has different performance requirements and cost constraints, organizations should select the most appropriate AI for each field rather than relying on a single model for every workload. In practice, low-cost models can handle routine customer inquiries while high-performance models take on complex programming assignments. Srinivas’ assessment is that selecting models according to the characteristics of each task is more rational in terms of both cost and performance than processing every request through a single top-performing model.

The shift in the AI market paradigm has been driven largely by “cost fatigue” among enterprise customers. The per-token cost incurred when using closed application programming interfaces (APIs) can vary by severalfold or even dozens of times depending on the workload and degree of optimization. As a result, a “hybrid strategy” that distributes open and closed models according to computational complexity is becoming standard practice across enterprises. For example, assume that processing the same task through a closed API costs $1.00 with 95% accuracy, while inference using an internally deployed open-weight model that a company can directly modify and operate costs $0.60 with 92% accuracy. If an orchestration system assigns 80% of routine work to the open model and routes only the remaining 20% of complex tasks to the closed model, the company can reduce its average cost by 32% while maintaining accuracy.

Peter Fenton, a general partner at US venture capital firm Benchmark, predicted that this transition could unfold faster than expected. He argued that the proliferation of open models could transform the very structure of the AI industry, as smaller models optimized for specific tasks are increasingly outperforming general-purpose large models in both speed and quality. “More than 90% of AI tokens generated over the next 18 to 24 months could come from open-weight models, and that could happen as early as the end of this year,” Fenton said.

AI developers including OpenAI, Meta and SpaceXAI have recently rolled out a succession of low-cost, high-efficiency models, intensifying competition over price-performance. OpenAI’s GPT-5.6 was designed to accomplish more tasks while consuming fewer tokens than its predecessors. SpaceXAI has also promoted its newly released Grok 4.5 model as delivering twice the token efficiency of comparable AI systems from rival companies. Meta’s new Muse Spark 1.1 model is likewise expected to focus on strengthening price competitiveness.

Changing Competitive Landscape Amid Proliferation of Open Models

The narrowing performance gap among frontier models lies behind the AI industry’s shift toward price-performance competition. AI competitiveness was previously evaluated largely on the basis of parameter counts, training scale and inference performance. The successive emergence of high-performance open-source models, however, is rapidly closing the performance gap. According to the AI benchmarking platform LMSYS Chatbot Arena, GPT-4o, Claude 3.5 Sonnet, Llama 3.1 405B and Gemini 1.5 Pro are separated by approximately 15 points. The industry generally regards differences of less than 50 points as difficult for users to perceive.

An analysis in the “2026 AI Index,” published by Stanford University’s Institute for Human-Centered Artificial Intelligence (HAI), reached a similar conclusion. As of March this year, Anthropic, SpaceXAI, Google and OpenAI were separated by no more than 25 Elo points on the Arena leaderboard, which incorporates human preferences. The gap between the top-performing US and Chinese models also narrowed to 2.7%.

Once quality reaches a certain threshold, the rising cost of achieving marginal performance gains begins to make price, latency, reliability and task suitability decisive factors in purchasing decisions. Notably, this shift is occurring even as the computational resources required for model development continue to increase. The AI Index found that while publicly disclosed parameter counts have remained at approximately 1 trillion over the past three years, the amount of computing power used for training has steadily increased.

As developers find it increasingly difficult to establish a performance lead substantial enough to overwhelm competitors despite committing enormous computing resources, the financial burden of developing large-scale models has grown further. The direct drivers of these mounting cost pressures are the proliferation of reasoning models and AI agents. According to an analysis of 100 trillion tokens by the AI model-routing platform OpenRouter and global venture capital firm Andreessen Horowitz, reasoning-specialized models accounted for a negligible share of processed tokens at the beginning of last year, but that figure rose to more than 50% by year-end. Average input length increased from approximately 1,500 tokens in early 2024 to more than 6,000 by the end of 2025, while average output length climbed from roughly 150 tokens to 400. Workflows that retain codebases, documents and conversation histories over extended periods while repeatedly calling tools and verifying results have driven up token bills.

The adverse effects of “Tokenmaxxing,” in which usage volume was treated as a proxy for productivity, have also surfaced across the corporate sector. Amazon recently discontinued an informal “KiroRank” system that ranked employees by AI token consumption and instructed staff to focus instead on solving customer and business problems. Uber similarly encouraged active AI use to the point of creating a token-usage leaderboard early this year, but it reportedly exhausted its entire 2026 budget for the technology in just four months after usage of Anthropic’s Claude Code surged. Insistence on using a single premium model caused service operating costs to rise exponentially. This explains why the early-stage strategy of maximizing AI usage is being replaced by “Valuemaxxing,” which measures output relative to cost.

Chinese AI Models Expanding Presence in Pricing and Usage

Efficiency-driven competition, however, is bound to reshape the AI rivalry between the United States and China. According to the AI Index, the performance gap between the two countries’ leading models has narrowed to 2.7%, while China has taken the lead in the number of AI research papers, citations and patent registrations. The United States released 59 notable models last year, surpassing China’s 35, but the performance gap among the highest-ranked models continues to shrink.

The shift is even more pronounced in pricing. Chinese AI company DeepSeek has priced its V4 Pro model at $0.435 per 1 million input tokens and $0.87 per 1 million output tokens. The price for cached input tokens falls as low as $0.003625. Compared with the $5 charged per 1 million input tokens for OpenAI’s GPT-5.6 Sol and Anthropic’s Opus 4.8, the difference exceeds tenfold.

This pricing advantage is translating into higher usage. According to OpenRouter, Chinese open models accounted for approximately 30% of total token usage during certain weeks last year, with an annual average share of 13%. Over the same period, the DeepSeek family processed 14.37 trillion tokens, while Alibaba’s Qwen family handled 5.59 trillion. Programming and technical workloads accounted for 39% of Chinese open-model usage. The figures indicate that inexpensive Chinese models are spreading into real-world production environments, including development, data processing and system operations.

China’s price competitiveness has strengthened as its companies adapted to US technology restrictions. With US export controls on advanced AI chips making high-performance accelerators more difficult to obtain, Chinese companies concentrated on developing technologies that could maximize throughput from limited computing resources. DeepSeek and Qwen, in particular, reduced memory consumption and computational burdens through techniques such as sparse activation, which selectively engages only the necessary parameters, and cache compression. They also lowered the cost and time required to migrate existing services to Chinese models by providing APIs compatible with OpenAI. Their ability to process more tasks with the same resources has translated into a pricing advantage, undermining the pricing power long enjoyed by US companies.

China’s abundant electricity supply is also supporting its aggressive pricing strategy. According to The Wall Street Journal, China generated more than twice as much electricity as the United States last year, while electricity prices at some Chinese data centers were less than half those paid by US facilities. The International Energy Agency (IEA) projects that global data-center electricity consumption will more than double from 415 terawatt-hours (TWh) in 2024 to 945 TWh in 2030. As demand for AI inference rapidly expands, electricity costs, generation capacity and the speed at which data centers can connect to the power grid are expected to become critical determinants of model pricing.

The United States, of course, still holds an advantage in capital, advanced semiconductors and cloud infrastructure. US private investment in AI is 23 times that of China, while the country has 5,427 data centers, more than any other nation. With Nvidia supplying more than 60% of global AI computing resources, US companies have also tightly integrated advanced accelerators, hyperscale cloud infrastructure, frontier models and enterprise customer networks. Yet massive investment and cutting-edge technology do not guarantee low inference costs. Experts warn that unless the United States converts its existing assets into lower costs per task and greater electricity supply, China could use the scale it has secured in high-volume processing markets to expand its influence further across developer ecosystems and enterprise markets.