AI is accelerating. It still will not ship a new model every day.
Samuel Hammond drew a straight line through model tenure and it reaches one new frontier model every 24 hours by January 6, 2027. Capability really is speeding up. That line is still the wrong shape for it, and the honest version of the trend is stranger than a crowded release calendar.
Three things are true at once. Capability is accelerating: Epoch's own breakpoint analysis finds the frontier improving about 1.85 times faster since April 2024, driven by reasoning models. Samuel's straight line is still the wrong shape, because it runs on tenure, which is floored by how fast labs can ship and dominated by GPT-4's long opening reign. And tenure is the wrong lens anyway. It measures how crowded the race is, not how fast the frontier moves. The likelier next step is not a model a day but models that learn continuously, where a discrete release stops being the unit. This uses real Epoch data. It is a reading of what has happened and where it points, not a dated forecast.
Capability is accelerating, and Epoch already measured it
Each dot is a model's Epoch Capabilities Index score the day it became the best in the world, from GPT-4 to today. Epoch's own breakpoint analysis, across 149 models, dates a speedup to April 2024 and puts it at about 1.85 times the earlier pace, with a 90 percent range of 1.3 to 3.1 times, driven by reasoning models. The flat stretch in 2023 is the year GPT-4 held the record with no challenger, not a year without progress. Read the magnitude from Epoch; this chart shows which models and which labs did the climbing.
The line, and why it breaks
Each bar is how many days a model held the top, the tenure Samuel regressed. His fit, R squared 0.34, reaches one new model every 24 hours by January 6, 2027. Two things break it. It leans on GPT-4's 352-day reign, the single longest, which pulls any straight line steep. And tenure is floored by how fast labs can physically ship, so a straight line has to level off well before a model a day. The 24-hour crossing is an artifact of the ruler.
Tenure measures the crowd, not the speed
Look at the record. GPT-4 held the top the longest, 352 days, and was also the single biggest jump on the index. The longest reign and the fastest leap were the same model. That is the tell: tenure does not track how fast the frontier moves. It tracks how many labs are close enough to trade the lead, which fell as OpenAI, Anthropic, and Google all reached the front.
The deeper problem with the forecast is that it assumes the discrete model release stays the unit. The likelier next step is the opposite: models that keep learning after they ship, updated continuously rather than replaced. In that world there is no new model every day, because there is no release day. Tenure stops meaning what the tweet assumes, and the thing to watch is the slope of the capability itself, not the calendar. That is our read of where this goes. It is a reading, not a dated forecast.
One model held the top for all of 2023. Since then a different model has taken it every six to eight weeks. That is the field filling in behind the frontier, not the frontier slowing down.
Real Epoch data, and where our confidence stops
Every point is a real Epoch Capabilities Index record-holder from GPT-4 to now, computed from Epoch AI's public scores and refreshed every day. A model counts when its ECI sets an all-time high on release. For the size of the acceleration we defer to Epoch's own breakpoint analysis rather than our record-holder slope. The dashed line on the tenure chart is reproduced from Samuel Hammond's published coefficients, so the overlay is his fit, not our restatement of it.
GPT-4 held the top the longest and opened the widest gap on everything before it. That single point pulls any straight line steep and does most of the work in an extrapolation like the tweet's.
We do not size the acceleration ourselves. Epoch's breakpoint fit over 149 models gives about 1.85 times faster since April 2024, with a 90 percent range of 1.3 to 3.1 times. Our record-holder chart shows the same shape with less rigor, so we cite theirs for the magnitude.
That discrete releases give way to models learning continuously is our read of where this points, not a measured trend and not a dated forecast. We flag it as a hypothesis, and the thing to watch is the capability slope itself.
The line is wrong. The pull under it is real.
AI is accelerating, but not toward a model a day. Toward models with no release day at all. Pass it on.