The Exponential Replacement Curve V2 and METR's November Acceleration Data

On November 5, METR published its assessment of AI research acceleration at Anthropic, based on a high-access evaluation that ran from August 12 to October 29. Its central estimate is that Anthropic's models are improving on well-specified tasks at about twice the rate they did between March 2024 and October 2025. On AECI, Anthropic's version of the Epoch Capabilities Index, the measured acceleration since October 2025 is 1.9x, with a 95% confidence interval of 1.5x to 2.8x. The index trend moved from roughly 15 points a year to roughly 29.
A chart that circulated after the report converts those AECI scores into task time horizons. In it, the doubling time for task length shortens from 3.3 months to 1.7 months, a ratio of about 1.9x that matches METR's headline figure. I used that chart to check the dates in the Exponential Replacement Curve V2
What the curve dates
Each V2 wave pairs a band of task lengths, measured in hours of human work, with the quarter in which models were projected to complete tasks of that length. The dates mark when a category of task becomes technically possible for a model to perform. They make no claim about when, or whether, an employer stops paying a person to do that work. That decision depends on cost, liability, regulation, and how the work is organized, and it can trail capability by years or never follow it at all. The curve tracks the capability event only.
The comparison
V2 set its thresholds on METR's 50% time horizon, the length of task a model completes half the time. The chart reports the 80% horizon. To get both readings, I converted with the chart's own ratio, dividing 50% hours by 6.3 to get 80% hours. Dates beyond about 38 hours assume the 1.7-month doubling continues, which is the chart's dashed extrapolation.
ERC V2 waveV2 horizonV2 capability dateChart, 50% basisChart, 80% basis1: Admin, customer service, contentup to 40hQ2–Q3 2025Apr–May 2026 (3 quarters late)Sep–Oct 2026 (5 quarters late)2: Sales, finance, software engineering40–640hQ1–Q3 2026Apr 2026 to Dec 2026 (starts on time, ends 1 quarter late)Oct 2026 to Apr 2027 (2–3 quarters late)3: Research, strategy, creative640–10,240hQ4 2026–Q3 2027Dec 2026 to Jun 2027 (on time to 1 quarter early)Apr 2027 to Nov 2027 (1–2 quarters late)
The chart starts behind V2 because V2's 2025 baseline was set too high, which put Wave 1 about three quarters early on the 50% basis. V2 assumed a 5.5-month doubling, and the chart runs at 1.7 months, so the two lines converge. On the 50% basis the chart meets the V2 timeline in Q4 2026 and is slightly ahead of it through 2027. On the 80% basis, which requires a model to succeed four times in five, the chart remains one to two quarters behind V2 through 2027.
Adjusting the chart
Two factors push the chart's dates later. Its data points are dated by when a model was available inside Anthropic, while V2 refers to capability available to the public, so by V2's definition the chart runs a few months early. METR's own figures show a similar gap: the Haiku 5.5 system card reported 1.7x acceleration, and METR attributes the difference from its 1.9x mainly to the inclusion of unreleased internal models. The fixed 6.3 ratio between the 50% and 80% horizons is also an assumption. The real gap varies by model, so the 80% column carries more uncertainty than the 50% column.
The data points are a third party's conversion of AECI scores into hours. METR's public report contains no time-horizon figures, and I could not confirm the 38-hour value.
What METR left open
METR found no clear evidence of continuing acceleration in AECI, though it said redacted evidence makes continued acceleration seem more plausible. Robustness checks across link functions, benchmarks, and model snapshots returned a range of 1.1x to 2.9x. The report raises the possibility that the gains are concentrated in well-specified, verifiable tasks, and it flags benchmark saturation: the model METR calls Model 4 scores above 50% on all 15 benchmarks behind the index. Anthropic redacted what METR described as a majority of the evidence it found most informative, and a compute-allocation verification step was canceled partway through the engagement.
The well-specified caveat applies directly to the curve's later waves. Much of the work in Waves 2 and 3, from financial analysis to strategy, has loose success criteria. If the acceleration holds mainly for tasks with clear criteria, my reading is that capability within those occupations would arrive task by task over a longer span than the wave dates suggest. Whether 1.7 months is a sustained doubling rate or a one-time step up has not been measured yet, and the next public models scored on METR's time-horizon suite will be the first direct test.
About the author
David Borish writes about frontier AI and its economic effects at davidborish.com. He is the author of The Tony Hawk Paradox: When Video Games Predict Reality tonyhawkparadox.com

