Self-reported coding uplift hit 2.5x at Opus 4.5, but large multiples are unreliable

maksym_andr · x · 2026-11-17

Continuing his thread on METR's independent assessment of Anthropic's AI R&D acceleration, maksymandr argues the self-reported uplift curve is more representative, capturing the step change from Opus 4 to Opus 4.5 (1.3x vs 2.5x). The chart of informal Anthropic employee surveys fits to roughly 3.9x/year exponential growth, with an alternative superexponential fit. He cautions that self-reported uplifts may be arbitrarily unreliable at large factors — judging 6x vs 12x is far harder than 1x vs 2x.

Related event: METR: AI Roughly Doubles Anthropic's Internal R&D Speed(3 posts)→

Original post →

More from Models

Models channel →