Anthropic models may be cheaper to use than some large open-source Chinese models, per a new study by AlphaSense.

It says Kimi is substantially cheaper per token than GPT-5.6 Sol, yet more expensive per completed question because it consumes many more tokens while assembling context.

The real unit of cost is not $/token, but tokens-to-completion × token price. A more expensive model that searches efficiently, stops at the right time, and reasons over a smaller, cleaner context can have a lower total inference bill.

"Token efficiency" is really a systems property, not an API-pricing property. Kimi K3 can have cheaper tokens yet cost more per question because poor stopping behavior and repeated retrieval inflate context before useful reasoning even begins; in multi-step agents, that waste propagates into later steps, so a small retrieval inefficiency can become a much larger workflow-level tax.

AlphaSense gets roughly a 3× cost reduction from changing retrieval around the same model, while task-level routing improves preference even further.

So that means model vendors can keep leapfrogging each other without fully commoditizing the application layer, because whoever owns retrieval, routing, and evaluation decides which model is economically viable for each task.