DEV Community

Cover image for Altman’s cheaper token paradox
Paulo Victor Leite Lima Gomes
Paulo Victor Leite Lima Gomes

Posted on

Altman’s cheaper token paradox

#samaltman #tokencost #token #ai

In February 2025, OpenAI CEO Sam Altman published a short blog post titled “Three Observations,” in which he argued that the cost of using a given level of AI falls about 10× every 12 months, dramatically faster than Moore’s Law (2× every 18 months). Yes, he said this, he pointed to the 150× drop in price per token from GPT-4 (early 2023) to GPT-4o (mid‑2024) as proof.

But just months after his post, a new trend: AI costs are no longer falling uniformly. While everyday models have become drastically cheaper, frontier models-GPT-4.5 and o1‑pro, have seen price hikes that break the previous curve. Lets fact‑checks the current situation and explains why both things can be true at once.

Sam Altman’s Prediction vs. Reality

reality

Altman’s observation was never that every model gets cheaper forever. He noted that for a fixed level of capability, costs plunge. However, when you demand higher intelligence, models that reason longer, handle more complex tasks, or achieve new benchmarks, the price can actually rise.

Since February 2025, OpenAI has released:

  • GPT‑4.5 (Feb 27, 2025) — a “research preview” with much higher intelligence, but also much higher compute cost.
  • o1‑pro (Mar 20, 2025) — an ultra‑reasoning model for specialists, priced at a premium.

quiet
Simultaneously, OpenAI quietly cut prices for GPT‑4o in mid‑2025, making the workhorse model cheaper than ever.

Fact-Check: Token Costs Over Time (2023–2025)

The table below shows official OpenAI API pricing (per million tokens) for key models, with percentage changes relative to the previous comparable model.

table-tokens

suspecious

  • Massive deflation for mainstream models, eg GPT‑4o mini is 96% cheaper than GPT‑3.5 Turbo
  • Steep inflation for frontier intelligence: o1‑pro costs 10× more per output token than the original GPT‑4
  • Percentage increase from GPT‑4o to GPT‑4.5: ~900% higher output price.
  • From GPT‑4.5 to o1‑pro: another +300%

Why the Divergence?

Two forces are at work:

Commoditization of “good enough” AI - GPT‑4o and mini benefit from engineering optimizations, distillation, and competition (from Google’s Gemini, Anthropic’s Claude, and DeepSeek). Their prices fall as predicted.

Scaling intelligence is expensive — GPT‑4.5 and o1‑pro use techniques like test‑time compute (the model “thinks” longer) and massive parameter counts. Each token requires far more FLOPs, and OpenAI prices accordingly.

As Altman himself wrote in “Three Observations” (without yet knowing the 2025 pricing): “The cost to use a given level of AI falls about 10x every 12 months” — but he also noted that the level of AI people want keeps rising. That demand for higher capability explains the premium tiers.

What does it means for developers and Businesses ?

For 95% of tasks (chatbots, summarization, classification, RAG), costs are plummeting. You can now run GPT‑4o for 2.50/2.50/10 per million tokens — cheaper than many specialized models from 2024.
For cutting‑edge reasoning (advanced math, code generation, multi‑step planning), you pay a luxury price. But those tasks were impossible with older models at any price.
In short: Moore’s Law for AI is alive and well for given level of intelligence. But if you want the next level of intelligence, be prepared to pay , at least until that level also gets commoditized in 12-18 months.

Also, I don’t like predictions, but we still need time to build a smart energy and servers to deal with this crazy demand. Maybe in 5 years or so.

References

Top comments (0)