Time Horizon Model
2026 VI / July calibration
Trajectory
50% and 80% task-success horizons, January through December 2026.
Monthly values
| Month | Released 50% TH | End 50% TH | End 80% TH | Uplift | Start difficulty | Start doubling |
|---|
Model explanation
The page uses a toy effective-scaling model inspired by Anthropic's Responsible Scaling Policy v3.1: compute progress and algorithmic efficiency multiply, so their rates add in log space. The May 1 observed doubling period calibrates the aggregate raw progress rate.
Algorithmic progress is lagged to completed models. A model completes on the first day of each month, and that released model sets the algorithmic term until the next month completes. Difficulty is evaluated continuously from the current frontier TH, so doubling time can change inside a month even when the released model is fixed. Effective difficulty is capped at 100%, so anchor extrapolation cannot make early dates harder than baseline.
The main growth model treats TH as the 50%-success horizon. The 80%-success horizon is shown as a proportional readout and second chart line: by default, 80% TH equals one-fifth of 50% TH, so May 1 26h 53m at 50% success corresponds to about 5h 23m at 80% success.
The Jan-Apr values are solved backward month by month. Given the May 1 TH, the model first finds the Apr 1 released TH that would grow to May 1 while using only the Apr 1 model for algorithmic progress. It then repeats the same calculation for Mar 1, Feb 1, and Jan 1.
The utility mode only changes the displayed uplift. Released TH power scales uplift with released model TH, total raw progress scales it with the full calibrated raw progress rate, and log raw progress uses a slower additive response to raw-progress multiples. Displayed uplift is capped at 500% in all modes.