AI inference costs drop roughly 13x a year, Epoch AI finds, outpacing prior technologies
Epoch AI measures about a 13x annual decline in the cost of hitting a fixed AI performance benchmark, with MIT attributing roughly 3x of that to algorithmic progress alone.
The cost of reaching a given level of AI performance is dropping faster than for any earlier technology on record, according to figures compiled by research group Epoch AI. Its measurement puts the annual decline at roughly 13x, meaning a task that cost a set amount a year ago can now be run for a small fraction of that. The finding covers a fixed performance target, not the price of any specific model on the market.
Two forces drive the number, and analysts separate them. Some of the decline comes from cheaper and faster hardware, and some from competition pushing prices down. MIT researchers strip those out and estimate that algorithmic progress alone accounts for about 3x per year — better methods and training recipes rather than better silicon. That residual figure is the one that speaks to how fast the field is genuinely getting more efficient per unit of capability.
The headline rate does not translate into cheaper bills for every user. Today's strongest models are not cheaper than their predecessors, because capability keeps moving the target. Reasoning models are the clearest example: they spend far more compute per task by generating long chains of intermediate steps, so a single query can cost more than a conventional model doing the same job. Falling unit costs and rising per-task compute are pulling in opposite directions.
That tension matters for anyone choosing a model for production use. Price is one input among several, and Epoch AI's benchmark framing is a reminder that cost curves describe the frontier, not a given deployment. Quality, latency, and error rate often decide whether a system is usable at all, and a model that is cheap per token but wrong more often can end up more expensive once retries and human review are counted.
The measurement also carries a caveat about how such curves are built. Fixed benchmarks age: once models clear a threshold, the cost of hitting it keeps falling even as the frontier moves to harder tests. Whether the 13x rate holds as evaluation shifts toward tasks that current systems still fail is the open question. Watch the next round of Epoch AI and MIT updates for whether the algorithmic share holds up — or whether hardware and price competition are doing more of the work than the headline suggests.