Berkeley and LSE Economists Estimate Markets Price In a 32.6% Software Engineering Productivity Gain From AI, Roughly Doubling by Mid-2026 as Coding Agents Spread
An NBER working paper infers investor expectations from stock moves: AI is priced as a permanent 32.6% software engineering productivity gain through 2025, with expectations more than doubling by mid-2026.
Overview
Economists at the University of California, Berkeley and the London School of Economics and Political Science (LSE) have estimated how much software engineering productivity gain stock markets are pricing in from AI, and they link a sharp rise in 2026 to coding agents. In a National Bureau of Economic Research (NBER) working paper dated September 2026, they say that from November 2022 to December 2025, AI increased the market’s expected present value of software engineering productivity by the equivalent of a permanent 32.6% productivity increase. The paper measures investor expectations, not developer output.
What We Know
- Authors and status. The paper, NBER Working Paper 35793, is by Alex Blumenfeld, Jonathon Hazell, Chen Lian and Andreas Schaab. Blumenfeld, Lian and Schaab are at the University of California, Berkeley; Hazell is at LSE, according to the paper. It states that NBER working papers are circulated for discussion and comment and have not been peer-reviewed.
- Headline estimates. The paper puts the corresponding effect on the level of GDP at 3.6% in the baseline and 6.5% when higher software engineering productivity also raises R&D productivity. It adds that by mid-2026, amid rapid progress in coding agents, the effect of AI on productivity and GDP had more than doubled relative to the end of 2025.
- Method. Rather than measuring developers’ output directly, the researchers examined how company stock returns respond to news about AI and whether that response varies with the proportion of each company’s payroll devoted to software engineering, as The Register reported. In the paper, the authors regress firm-level stock returns on the return of the ROBO Global Artificial Intelligence index (THNQ), and take software engineering payroll shares from Revelio Labs’ employee-level data. Their second-stage regression yields an estimated slope of 1.24, according to the paper.
- Sample. To isolate productivity, the authors exclude firms in software-producing industries or the semiconductor supply chain. Firms that use software engineering intensively in the sample include eBay, Sonos, Expedia and Airbnb, per the paper.
- 2026 acceleration. The authors write that news about AI-driven software engineering productivity in the first half of 2026 exceeded the total over the previous three years, and that one likely reason is the widespread introduction and adoption of coding agents such as Claude Code and Codex in 2026, according to the paper. The paper’s 3.61% GDP figure for November 2022 to December 2025 excludes this later news, the authors say in the same paper.
- Comparison with experiments. The authors say their estimate is comparable in magnitude to the 21-56% task-level speed-ups found in the experimental literature, in the paper. They also write that complementarities and bottlenecks between tasks imply aggregate productivity gains should be smaller than task-level gains.
What We Don’t Know
- Whether the gains materialize. The authors write that the view from financial markets is unlikely to be perfectly accurate, in the paper. Lian told The Register that markets can be overly optimistic or pessimistic, as reported by The Register.
- Employment effects. Lian said a preliminary look at the data suggests total software engineering employment among the covered firms has increased over the past few years, but that the productivity estimate does not depend on that employment trend, according to The Register.
- Agent-specific effects. The paper names coding agents as one likely reason for the 2026 jump, not as a measured cause; a stock-index method does not isolate any individual product.
- Other sectors. Lian said the methods can be used to study the economic impact of AI through other channels, which the team plans to explore in follow-up work, per The Register.
Analysis
The study sits between two kinds of evidence on AI coding tools: controlled task-level experiments and company-level adoption reports. By reading expectations from stock prices, it offers a real-time indicator, but one that reflects what investors believe rather than what teams ship. The authors say they plan to monitor the series as the frontier of AI continues to expand.