U.S. AI Models Beat Cheaper Chinese Rivals on Total Task Cost

Generative AI buyers cannot judge economics from the price of one million tokens alone. The number of reasoning steps, output length, retries and answer quality can materially alter the bill for completing a task. AlphaSense’s analysis shifts the comparison to total execution cost, weighing Chinese open-source models against proprietary systems from U.S. developers OpenAI and Anthropic as enterprises face pressure to prove returns on rising AI spending.

AlphaSense’s 2026 data put GPT-5.5 at $30 per million output tokens, versus $0.87 for DeepSeek. Yet its tests found U.S. models could finish work with fewer tokens and deliver higher-quality results, making the overall task cheaper despite a steeper list price. Token consumption in agentic workflows has risen as much as 30-fold, while Anthropic moved its enterprise tier to usage-based billing in April, strengthening the case for routing each workload to the most cost-effective model.