The length of programming tasks OpenAI's reasoning models have been capable of has been doubling every three months. This is likely to continue until the scaling of reasoning models hits the compute ceiling, likely sometime in 2026, at which point the doubling-time may revert to something just a little better than the previously-observed doubling time of 7 months.

The period of time between the release of o3 and when we expect the 3-month doubling period to end is 13 months. This entials 13/3 = 4 1/3 doublings, which would get us to 1935 minutes, or 32.25 hours. We can then expect one more double, to 64.5 hours.

So, by EOY 2026 I expect we'll have coding agents that can do the sorts of things that would take a skilled human engineer a week of total focus to accomplish. Should be fun!