AI benchmark performance costs dropping 13x per year says Epoch AI
1 source · The Decoder- Boom: Epoch AI measures AI benchmark performance costs falling roughly 13x per year
- Boom: MIT isolates algorithmic progress alone at approximately 3x annual improvement
- Doom: Reasoning models can cost more per task due to higher compute consumption
- Neutral: Cost declines apply to fixed benchmarks, not necessarily to the latest frontier models
The story in full
Epoch AI has measured AI performance costs falling at roughly 13 times per year, a rate described as faster than any previous technology. MIT researchers, stripping out hardware improvements and market competition, put annual algorithmic progress alone at approximately 3 times per year.
The cost decline applies to a fixed benchmark performance level, not to frontier models overall. Reasoning models can cost more per task because they consume significantly more compute. Epoch AI and MIT attribute the gains to separate factors: hardware advances and competitive pressure on one side, and algorithmic improvements on the other.
Analysis
357 wordsEpoch AI published measurements showing that the cost of reaching a fixed AI benchmark performance level has been falling at roughly 13 times per year, a rate the organization describes as faster than any comparable technology in history. Separately, MIT researchers isolated the contribution of algorithmic improvements alone, stripping out hardware advances and competitive market pressure, and found that factor accounts for approximately 3 times annual cost reduction on its own. Both findings were reported in late September 2026.
The distinction between the two figures matters considerably. The 13x figure captures the combined effect of better chips, lower prices driven by competition among cloud and model providers, and more efficient algorithms. The 3x figure is narrower and arguably more durable, since algorithmic progress tends to compound regardless of market conditions. Critically, neither figure means that using the most capable frontier models is getting cheaper. Reasoning models, which chain together many inference steps to tackle harder problems, can consume far more compute per task than earlier systems did, which pushes their per-task costs upward even as benchmark costs fall. The cost declines apply to a fixed performance target, not to whatever the leading edge happens to be at any given moment.
No reactions from the Pro-AI, Anti-AI, or Middle Ground camps have been published yet on this story. The Pro-AI camp would typically treat findings like these as confirmation that AI capabilities are on a trajectory to become broadly accessible, with falling costs accelerating adoption across industries. The Anti-AI camp would likely focus on the caveat about reasoning models, arguing that real-world deployment costs are rising rather than falling and that headline efficiency numbers obscure the resource intensity of cutting-edge use cases. The Middle Ground camp would probably accept both data points as valid, emphasising that the benchmark framing requires careful interpretation before drawing conclusions about practical affordability.
The argument is likely to sharpen as reasoning models become more central to commercial deployment. Tracking how per-task costs for those models evolve over the next year, alongside any updated Epoch AI or MIT measurements, would offer the clearest evidence for which interpretation of the cost trend holds in practice.
Where do you stand?
Add your take
0 reader votesSign in with Google to pick a side and post. Your vote moves the story's Doom / Boom score.
