How fast is AI inference actually becoming more efficient?

To separate pure technical progress from market and hardware effects inside the broader 5×–10× annual price-performance decline, the paper runs its price regression on open-weight models only, "divide[s] the estimated price trends by the decline in hardware prices as reported in Rahman (2024)," and reports that "after these adjustments, the remaining price reduction factor is around 3× per year, which can be interpreted as the contribution of algorithmic progress to declining price." The reasoning for isolating open-weight models is that "open-weight models can, in principle, be run and modified by anyone," so "we expect that they will not be priced at a significant markup relative to the necessary GPU resources to run them" — making their price trend a cleaner proxy for technical progress than proprietary pricing, which can also move with competitive pressure.

The decomposition is supported by a comparison across model tiers: the closed-weight price trend runs slightly faster than the open-weight trend overall, and "particularly pronounced for closed-weight models in the 40%–60% group," where "a sudden drop in price... is not mirrored in open-weight models, hinting at non-technical competitive effects." The paper reports the resulting ~3×-per-year algorithmic estimate as "more similar to experimental measures of energy efficiency gains like (Saad-Falcon et al., 2025) which find 3.1× gains from 2023-2025 controlling hardware" than to the larger headline price-decline figures, framing it as convergent evidence from a different measurement method rather than a single-study estimate.

This decomposition supplies the component breakdown behind Is benchmark progress outpacing the cost to achieve it?: of the 5×–10× annual price-performance improvement, roughly 3× per year is attributed to algorithmic progress once hardware and open-vs-closed competition are accounted for, leaving the remainder to hardware efficiency and competitive pricing. The move — splitting one aggregate annual rate into separate multiplicative drivers — parallels Are AI feedback loops strong enough to sustain recursive self-improvement?, which likewise decomposes a single headline acceleration figure into distinct component elasticities rather than treating it as one undifferentiated trend.

The 3×-per-year algorithmic estimate depends on open-weight models being a representative stand-in for frontier technical progress with only "a lag of a few months to a year" (citing Denain, 2024), and on Rahman (2024)'s hardware price series being an accurate deflator; the paper does not independently validate either assumption within this analysis. It also does not estimate how the algorithmic, hardware, and competition shares might shift as open-weight and closed-weight ecosystems diverge further, so the ~3× figure should be read as a point estimate for April 2024–November 2025 rather than a stable long-run constant.