Not surprisingly, the cheap AI models are starting to win favor.

The collapse in performance between best and second best, low switching cost, and the plummeting cost of training new AI models are driving the market towards cheaper alternatives.

Performance: “Using the Epoch Index as a gauge, the leading AI product is now just 6% more capable than the one at the 90th percentile, an ‘almost invisible’ difference…”

Cost: “GPT-5.5 and Claude Opus each charge $5 for a million tokens, compared with just $0.019 for DeepSeek.”

Outcome: “The Information reported on June 23, a service that compares AI models so developers know which ones should power their apps or handle different tasks within them, show that the share of all tokens processed on open-source models—primarily DeepSeek and MiniMax from China—climbed to roughly 65% in the second week of June, up from 34% during the same week in January and 26% the year prior.”

Per the article, Lindy AI switched from Claude to Deepseek, Coinbase Global, Inc. has announced a goal to move 80% of its workloads to low-priced alternatives by end of year, and even MSFT is considering adding DeepSeek to its product lineup.

AI
Source: Evan Lorenz | "AI Outsmarts Itself" | Grant's Interest Rate Observer | 07/03/2026 | Visit