Only about 1.5 years have passed since the May 17, 2024 prediction, which explicitly talks about an ongoing trend of models becoming ~90% cheaper per year and being ~99% cheaper over any two‑year span. The first full two‑year window (to compare “models two years prior”) would end around May 2026, which hasn’t occurred yet.

We do see substantial price drops and efficiency gains around the time of, and shortly after, the prediction:

• Between the original GPT‑4 (March 2023) at $30 per million input tokens for the 8k context model and GPT‑4‑Turbo (April 2024) at $10 per million, OpenAI cut the per‑token price by about 3× in a bit over a year, a ~67% cost reduction but not a full 90%. (help.openai.com)
• GPT‑4o, launched May 2024, was announced as delivering GPT‑4‑level intelligence at half the cost and twice the speed of GPT‑4‑Turbo, a further ~2× drop in price at similar or better quality, but again not a clean 10× year‑over‑year step. (investopedia.com)
• GPT‑4o mini (July 2024) is priced at $0.15 per million input tokens and $0.60 per million output tokens—roughly two orders of magnitude cheaper than the original GPT‑4 pricing for frontier‑level models—and OpenAI describes it as “an order of magnitude more affordable than previous frontier models” while often outperforming GPT‑4 in chat preferences. (openai.com)
• In 2025, newer models like GPT‑4.1 and o3‑mini are described as cheaper than their immediate predecessors (e.g., GPT‑4.1 about 26% cheaper than GPT‑4o; o3‑mini about 63% cheaper than o1‑mini), continuing the downward trend but at increments far short of a consistent 10× drop each year. (theverge.com)

However, the prediction is about future sustained rates (“will continue to become about 90% cheaper every year … every two years these things are gonna be 99% cheaper and better”), not just the one‑off transition up to GPT‑4o. We have:

• Less than the two years he explicitly uses as the benchmark window.
• No standardized, industry‑wide metric for “cost to run” across all “AI foundation models,” since provider prices mix model quality, context length, and product strategy.

Because (a) the forecast period he specifies has not fully elapsed, and (b) existing data, while showing large cost reductions, is not sufficient to confirm or falsify a continuing ~90%‑per‑year drop, the status of this prediction as of November 30, 2025 is inconclusive (too early).