The best AI in the world is about to change every day.
Then every hour. Then every minute.
For most of a year, GPT-4 stood alone at the top. Now the best model in the world is dethroned every few weeks, and the gap between one champion and the next keeps collapsing. Draw the trend straight and it reaches a new frontier model every 24 hours by January 2027, then every hour, then every minute.
Right now: about one every 16 days. And still accelerating.
Each dot is a model that was, briefly, the best in the world. Its height is how many days it held the top before something beat it. Grab the blue line and tilt it to steepen or soften the trend, and watch the date move.
GPT-4 stayed on top for 236 days. Claude Fable 5 managed thirty. Getting from GPT-3 to GPT-4 took nearly three years; the best model in the world now changes in weeks.
Try a different line
pick one, or drag the line yourselfSamuel Hammond's straight line. It reaches one a day by Jan 6 2027.
Soon, a new best model every…
Read off the current line, the date each pace arrives. Every step lands sooner than the one before it.
- every monthMar 14, 2026the trend already passed this
- every weekNov 5, 2026one every seven days
- every dayJan 6, 2027the headline
- faster stillJan 15, 2027an hour, then a minute, then a second, all within the same few days
The pace from here
How often a new best model would land, reading the current line forward.
Why this is probably wrong, on purpose
A straight line through a noisy trend tells you something is speeding up. It does not tell you a new model will literally ship every 24 hours on a given day. The point is the shape of the curve, not the calendar. The honest caveats below are what make the trend worth taking seriously in the first place.
GPT-4 stayed on top far longer than anything since, so it pulls the line steep. Drop it and the one a day date slips well past 2027. Try the switch.
A pace like this is more likely to bend than to fall off a cliff. A smooth curve eases toward zero and never quite reaches it, which pushes the date out by years.
Evaluations, safety review, and shipping do not compress the way training does. And models may stop arriving as discrete versions at all, and simply keep updating.
The math, kept honest: Samuel's line is tenure ≈ 136.7 minus 35.6 times the years since GPT-4. It explains about a third of the scatter (R squared 0.34, n = 18). We pin that line so it reproduces his Jan 6 2027 date exactly. The other three switches refit live from the data, so logging a new model updates them on their own. The dots are the real ECI record-holders from Epoch's public scores, which is why fitting a line to them lands right on his numbers.
Where does the line bend?
Quote the chart with your own call on how fast this actually goes.