AI's Task Horizon Now Doubles Every 4 Months — Down From 7
The cadence of progress is itself accelerating: the time an AI can work autonomously is doubling every 4 months instead of 7. The reason is uncomfortable — Claude is now building Claude.
NeuroX AI · June 10, 2026
Anthropic just published the loop everyone suspected was forming. The length of task an AI can complete autonomously is doubling every 4 months — down from a 7-month doubling rate a year ago. The curve isn't just steep. Its slope is increasing.
The mechanism is recursive. Claude Opus 3 (March 2024) handled ~4-minute tasks; Sonnet 3.7 reached ~90 minutes; Opus 4.6 now sustains 12-hour runs. That acceleration tracks a second number: 80% of the code merged at Anthropic is now authored by Claude. The model is shortening the cycle that produces the next model.
The internal evals are the part to sit with. Claude's research-direction judgment went from beating a human's choice 51% of the time (Nov 2025) to 64% (April 2026). On code optimization, the speedup over a human baseline climbed from ~3x in May 2025 to ~52x by April 2026. Human researchers hit roughly 4x on the same tasks.
Here's the operational read for everyone not at a frontier lab: capability is compounding faster than most teams' verification layer. A 4-month doubling means the agent you scoped around in Q1 is a different tool by Q3. Pipelines built for a fixed ceiling age out before they ship.
The advantage isn't access to the model. It's a review-and-CI layer that absorbs a moving one.