# AI inference costs falling 1000× in 2 years, heading toward free intelligence

Source: https://farcaster.xyz/kazani/0xd740a14d
Kiwi News: https://news.kiwistand.com/stories/AI-inference-costs-falling-1000-in-2-years-heading-toward-free-intelligence?index=0x693fde0abeae73196b3e495e500462504ec5f8e16902377d3fb5022d760d203754080919
Submitted by @kazani on 2025-12-15 UTC, 3 upvotes, 0 comments

## Summary

Falling AI inference costs suggest intelligence may soon become too cheap to meter, the text claims. GPT-4-class inference prices reportedly fell about 1000× over roughly two years. ARC Prize reports about a 390× drop in cost per task for near-state-of-the-art ARC-AGI-1 performance over about a year, from o3 preview at roughly 88% and an estimated $4,500 per task to GPT-5.2 Pro at 90.5% and $11.64 per task. If these trends continue, the marginal cost of white-collar cognition could become orders of magnitude cheaper than human labor once human-level AI is reached.

## Article

We’re heading towards a future in which intelligence is too cheap to meter.

1. “GPT-4-class” inference prices have fallen by ~1000× over ~2 years.

2. ARC Prize reports ~390× lower cost per task for near-SOTA ARC-AGI-1 performance over ~1 year (o3 preview: ~88% at est. ~$4.5k/task → GPT-5.2 Pro: 90.5% at $11.64/task).

3. GPT-5.1 sometimes needs up to ~2× more output tokens than GPT-5.2 for similar SWE-Bench performance; and on ARC-AGI-2 up to ~2.7× more tokens for similar performance.

If these trends continue, then once we reach human-level AI, the marginal cost of “white-collar cognition” could rapidly become orders of magnitude cheaper than human labor.

Sources: 

- https://andreyfradkin.com/assets/LLM_Demand_12_12_2025.pdf
- https://x.com/arcprize/status/1999182732845547795
- https://x.com/tobyordoxford/status/1999870642032967987