AI thought price falls 13x per year, Epoch AI finds — 725x drop in 18 months
Epoch AI · rss · 2026-09-28
Epoch AI published a summary of its report on how fast the "price of thought" — the cheapest cost of hitting a given benchmark score — is falling:
- Over five years, cost per performance level has dropped 13x per year (47% per quarter): 4x faster than DNA sequencing, 6x faster than compute, 18x faster than lithium batteries, 54x faster than electricity.
- Example: in January 2025, o3 hit 75% on GPQA Diamond at $0.30/question; under 18 months later GPT-5.6 Luna matched it at $0.0004/question — a 725-fold drop, like a $50,000 car falling to $69.
- By domain: math prices fall fastest (50–52% per quarter, 16–19x/year), game-based puzzles slower (39–43% per quarter).
- The rate of decline slows with age: 66% per quarter when a level first achieved, halving to 32% two years later — likely a brief premium before competition drives prices down.
- The trend traces back to commercial LLM inference in November 2021; data and code are on GitHub.
More from Infra
- Sarvam AI Launches Token Factory: India-Hosted Inference for Open Models — itsOmSarraf_ · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Stratechery's Fictional Obituary: Imagining Jensen Huang's 2026 Exit From Nvidia — Gregory_C_Allen · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16