How fast is AI?
Menu

· monthly note

November update: the price of thought and the gap between adoption and wages

New capability scores push the frontier to ECI 169, open-weight models cross 158, and administrative labor studies show why rapid workplace adoption still moves wages and hours slowly.

On the fast clock, Epoch AI's latest capability fits move the frontier to 169.1 (GPT-6 Astra, followed closely by Claude Opus 5.5 at 168.4 and GPT-6.1 Terra at 168.1). Open-weight releases continue to close the distance: DeepSeek V4.1 Flash (ECI 158.8), Meta's Muse Spark Open (155.1), and Reflection's Beam 1.0 (150.2) all scored at or above the original GPT-5 mark, keeping the open-weight lag around four and a half months.

A common objection to token-price trackers is that reasoning models spend many more tokens per answer, making per-token price drops look steeper than the true cost per task. Luke Emberson and David Roodman's September 2026 Epoch AI study, 'The Plunging Price of Thought', tests this directly across five math, science, and puzzle benchmarks. Even counting full reasoning token spend per question, the cost of a fixed level of performance has fallen about 47 percent per quarter (13 times a year) since 2023, dropping fastest right after a capability level debuts (75 times a year) before slowing to 4.7 times a year two years later. We have added a reasoning-workload toggle to the intelligence ladder so you can compare standard chat and reasoning-heavy token blends directly.

On the slow clock, administrative and survey studies help explain why fast tool uptake has not yet bent aggregate employment or productivity curves. In the United States, Alexander Bick, Adam Blandin, and David Deming find that workplace adoption of generative AI outpaced both the personal computer and the internet at the same stage after launch, assisting 1 to 5 percent of work hours with self-reported time savings around 1.4 percent of total hours.

In Denmark, Anders Humlum and Emilie Vestergaard link adoption surveys of 25,000 workers and 7,000 workplaces across 11 exposed occupations to matched employer-employee tax and hours records. Although chatbot adoption and employer training are widespread, with workers reporting average time savings around 3 percent, difference-in-differences estimates find precise zero effects on earnings and recorded hours two years on, ruling out shifts larger than 2 percent.

Taken together with Fabrizio Dell'Acqua and co-authors' field experiment on the 'jagged frontier' and Kristina McElheran and co-authors' Census study of industrial AI adoption, the pattern across countries remains consistent with the weak-link view: workers adopt fast on individual tasks, but turning 1 to 3 percent task time savings into firm-wide output and pay takes workflow redesign, verification, and complementary investment. Reality remains on the baseline path.

Source: Linked sources distinguish commentary from research; results retain their study-specific scope. · Emberson & Roodman (Epoch AI, 2026): The Plunging Price of Thought (benchmark cost study) · Humlum & Vestergaard (NBER WP 33777): Large Language Models, Small Labor Market Effects (Danish admin data) · Bick, Blandin & Deming (NBER WP 32966 / Management Science 2026): The Rapid Adoption of Generative AI (US survey) · Dell'Acqua et al. (HBS WP 24-013): Navigating the Jagged Technological Frontier (field experiment) · McElheran, Yang, Brynjolfsson & Kroff (AEA P&P 2026): The Adoption of Industrial AI in America (Census firm data)