Cheap AI Can Cost More

Chinese AI models have rapidly gained attention for offering powerful capabilities at substantially lower token prices than leading U.S. models. But new research from AI market-intelligence platform AlphaSense suggests that lower token prices do not necessarily mean lower enterprise AI costs.

AlphaSense evaluated nine AI models across roughly 245 financial-analysis tasks involving earnings transcripts, SEC filings, analyst estimates and acquisition activity. Its comparison included OpenAI’s GPT-5.6 Sol, Anthropic’s Opus 4.8, Moonshot’s Kimi K3 and Zhipu/Z.ai’s GLM-5.2.

The results challenge the conventional pricing argument. Despite Kimi K3’s lower advertised output-token price, AlphaSense found GPT-5.6 Sol produced approximately 20% higher-quality answers at about 13% lower overall cost. Opus 4.8 reportedly delivered around 13% higher quality at roughly half Kimi K3’s total cost.

The reason is efficiency. More capable models may complete complex tasks with fewer tokens, reasoning steps and follow-up prompts. For enterprises, therefore, the meaningful metric is not simply cost per million tokens, but cost per successfully completed task.

However, other research presents a different picture. Artificial Analysis reportedly found Chinese models considerably cheaper under its methodology, while OpenRouter data has indicated major cost advantages for open-source Chinese models across broader workloads. This suggests that economics can change significantly depending on task complexity, quality requirements and measurement methodology.

The emerging lesson is that enterprises should avoid choosing between “cheap Chinese” and “premium Western” models as a blanket strategy. Routine summarization, classification and other commodity workloads may justify lower-cost models, while complex financial, legal or analytical tasks could favor more capable models if they produce accurate results with fewer iterations.

The future therefore points toward intelligent model routing. Enterprises can benchmark models against their own workloads and dynamically send each task to the lowest-cost model capable of meeting the required quality threshold. As competition drives prices lower across the industry, quality-per-dollar and cost-per-outcome—not token price—could become the real measures of AI economics.