Epoch AI: cost of a fixed level of AI performance drops 47% per quarter, 725-fold on GPQA Diamond in 18 months
- Epoch AI's report "The Plunging Price of Thought" (Emberson and Roodman) finds the cost of a given level of AI performance fell about 47% per quarter over the past three years, a 13-fold drop every year and, the authors say, faster than any other transformative technology in history.
- OpenAI o3 paid about $0.30 per question to reach 75% on GPQA Diamond in January 2025, while GPT-5.6 Luna reached roughly the same score for $0.0004 in mid-2026, a 725-fold decline in under 18 months.
- Tabarrok argues the drop in inference cost at fixed performance shrinks the open-model threat to frontier labs, since frontier models are not only better but also rapidly cheaper to run at any given performance level.
- The trend is not only smarter models: a given level of intelligence now needs dramatically less inference expenditure, so performance and cost are improving together.
Hacker News opinions
The price of human intelligence is falling too. I suspect looks are about to go up in value a lot faster.
Counterpoint: the price of human intelligence is actually rising, because there's ever less of it and we invented a cheap substitute that makes people stupider.
I've worked alongside physics PhDs, fluid dynamics and turbine engineers, chemists, and geology postdocs while I was a self-taught programmer making websites. There's a huge oversupply of smart people that software dev was soaking up. If SDE collapses as a field, a lot of them are in for a shock.
People ask me how you get social mobility in a post-AGI society where the value of intelligence and labor is gone. I unironically answer 'be hot'.
If intelligence stops being the trait that moves people up the socioeconomic ladder, looks won't replace it. The ability to project physical force becomes currency, and that's how civilized societies could completely unwind.
Where exactly is human intelligence getting cheaper, and 'rapidly'? I can see an argument for a decline in a global stagflation, but that has nothing to do with AI as far as I can see.
LLMs still need material for self-education, so somebody has to be intelligent enough to produce it. Otherwise their 'intellect' freezes at some point.
Why would it freeze? They can just generate more training material, and synthetic training data has been shown to be just as good if not better.
That's not necessarily true. Think of how many problems in mathematics are much easier to pose and to check than to solve.
They can just train on tasks solved verifiably: reinforcement learning from verifiable rewards. Pure next-token supervision is only the pretraining phase now, and there's no obvious taper-off at human level.
The report comes from Epoch AI, which is financed by funds like Schmidt Sciences, and Tabarrok praises it uncritically. Maybe what's 'falling rapidly' is human intelligence or moral standards.
What ideological message could that blog post be pushing? The report is about AI performance and cost.
Nobody will ever be accountable for these LLM systems' stability or safety. The data has to be supervised by a live human who understands responsibility and can actually be held to it.
When intelligence is treated as a means to an end, of course its 'price' falls. What's being bought is artificial intelligence, the price of real intelligence is unchanged, and what's actually dropping is the standard of living.
I disagree there: there's no such thing as artificial intelligence, there's only intelligence. And intelligence is a means to an end regardless of whatever else it does, it's a dimension of the survival problem.
This is exactly what the property-owning class wants. Smart, educated people's time being expensive has always galled them, and capitalism just found a way to turn intelligence into a purchasable commodity too.
The falling price might hold for text outputs, but not for other modalities. Image, speech, voice, and embedding models: embedding models haven't even been updated in years.