A 12% AI Work Score Is Not a Forecast of Jobs Replaced

The leading model on GDPx, a new test of demanding professional work, scores 12%. That sounds easy to interpret: perhaps AI can do about one job in eight. But the number does not count jobs, workers or even simply the fraction of tasks completed.

SemiAnalysis introduced GDPx on November 2, 2026. Its current version combines 206 tasks from 14 partners across 11 domains. On the leaderboard checked November 6, Claude Opus 5.5 leads at 12.0%. The useful question is what went into that percentage before it reached the scoreboard.

GDPx is designed to expose difficult work that remains beyond today's agents. Its score combines different grading methods and gives each domain equal weight. That can make a demanding evaluation useful without turning it into a forecast of employment.

Different Grades Answer Different Questions

Some GDPx tasks use an all-or-nothing grade. Others award a continuous score. The headline measure uses each task's intended method and averages three attempts; a separate “Partial” view is not the same headline score.

Consider an invented example, not a reconstruction of GDPx. An assistant tackles two assignments with ten equally weighted requirements each. It satisfies nine requirements on both.

If we count requirements, its average is 90%. If each assignment is accepted only when all ten requirements are met, its completion rate is zero. The underlying work has not changed. The question asked of it has.

Neither number is automatically dishonest. Partial credit can reveal progress that an acceptance test hides. A strict pass can capture something partial credit overlooks: one missing requirement may make the result unusable. A buyer needs to know which requirement failed, not merely how close the score came to 100.

GDPx's published software example makes that distinction concrete. In a Git-LFS task, an agent must fix commands that encounter missing historical file contents while preserving the behavior of a complete repository and meeting other constraints. Fixing the obvious error is only part of the assignment. The creators' account reports no complete pass in the selected attempts it analyzes. That is evidence about this task, not a failure rate for software engineering generally.

Equal Domains Are Not Equal Shares of the Economy

The headline score also averages across domains, with each receiving the same weight. Where a domain has several partners, those partners are weighted equally too.

This means a domain does not get a larger share of the final score simply because it supplies more tasks. Nor does it gain weight because it employs more people or accounts for more national income. The benchmark's creators deliberately chose fields they expect to benefit strongly from frontier AI, rather than building a representative sample of occupations.

That choice makes the score a balanced comparison across their selected fields. It does not measure how a random working day is spent.

The gap matters in both directions. A low result on exceptionally difficult assignments does not show that AI has little value as a supervised assistant. A higher result would not establish how many employees could be removed from a business. Jobs combine tasks, responsibilities and interactions; a benchmark grade does not specify how those pieces fit together.

The Tools Are Part of the Result

GDPx generally uses each model's native agent software when available, and OpenCode otherwise. Its default task limit is eight hours, with shorter limits for some tasks. The result therefore belongs to a model working within a particular setup, not to model weights in isolation. Our explanation of coding-agent test setups shows why that distinction matters.

There is also an audit boundary. GDPx's public rollout collection makes agent traces from one task per partner public; most tasks remain private. The creators' evaluation is not an independently reproduced test of every workplace.

The useful next question is not whether 12% sounds impressive or disappointing. It is whether the relevant task resembles the work being delegated, whether its grade captures an acceptable result, and what still requires a person. A benchmark can make unfinished work visible. Turning its score into a jobs forecast would hide the very distinctions that make it informative.