Updated September 2026.
GDPval is an economic evaluation framework developed by OpenAI to judge whether AI systems create measurable value in real work, not how intelligent a model appears in a benchmark. It asks a practical question: does using AI reduce the cost of producing work that a professional would actually accept?
That question matters directly for industrial AI. A cost estimate, a process suggestion, or a routing proposal from an AI system has very little value if it only looks impressive. It has real value when an estimator or manufacturing engineer can accept it, correct it where needed, and use it faster than starting the calculation from scratch.
Many AI pilots across industries never make it past the demo stage, in part because they are evaluated on the wrong thing: how capable the model looks, rather than how much of its output a domain expert is willing to accept and rely on.
From intelligence to accepted output
GDPval does not ask whether a model is intelligent. It asks whether it can produce a deliverable that an experienced professional would prefer over, or accept in place of, their own work. In the original study, tasks were drawn from real work performed by industry professionals with an average of fourteen years of experience, across forty four occupations, and judged through pairwise expert comparison: would the expert prefer the AI-generated deliverable, or the human one?
That is the same standard we apply to Blake, our manufacturing estimating and quoting product. An AI-generated cycle time, cost breakdown, or process suggestion is only useful if an estimator would accept it, or correct it quickly, rather than redo the calculation from scratch.
Why the economics are not linear
The GDPval paper frames the workflow most organizations already use in practice:
AI produces an initial output.
An expert reviews it.
Only if the output is rejected does the expert redo the task from scratch.
The expected cost of that workflow can be written as:
Expected cost (EC) = Model cost (MC) + Review cost (RC) + (1 minus w) x Human cost (HC)
Here, w is not accuracy. It is the probability that the AI output is accepted.
This is why AI value is non-linear. Acceptance determines how often the organization still has to pay the full cost of starting over. AI speed and review cost stay roughly constant, so every increase in acceptance rate reduces the expected cost faster than the acceptance rate itself increases.
A simple example
Consider a single, repeatable estimating task.
A manufacturing engineer completing the calculation from scratch costs $1,000 in time.
An AI system generates a first estimate for $10, which the engineer reviews for $40. The AI-assisted path costs $50 upfront.
If the AI estimate is not acceptable, the engineer still has to complete the calculation from scratch. So:
Expected cost (EC) = $50 + (1 minus w) x $1,000
At 50% acceptance: $50 + 0.50 x $1,000 = $550.
At 60% acceptance: $50 + 0.40 x $1,000 = $450.
At 70% acceptance: $50 + 0.30 x $1,000 = $350.
The cost ratio between human-only execution and AI-assisted execution (HC / EC) moves from roughly 1.8x at 50% acceptance to nearly 2.9x at 70%. Each step of acceptance improves the underlying quality by the same amount, but the economic advantage accelerates, because the most expensive path, starting over, is avoided more often.
Why we build around accept, correct, use
This is the reasoning behind how Blake is built: AI proposes a calculation, an estimator reviews it, and the accepted or corrected result becomes the quote. The value is not in the AI looking capable. It is in how often, and how quickly, a real estimator can move from proposal to a number they are willing to stand behind, and in what that correction teaches the system for next time.
Subduxion is building a world model for precision manufacturing. Our research focuses on how geometry, requirements, manufacturing processes and production outcomes can be represented and learned computationally, measured the way GDPval measures value: by what a professional actually accepts and uses. Read more about the product this thinking shapes in Blake, or about our broader research in research.
Reference
Patwardhan, T., Dias, R., Proehl, E., Kim, G., Wang, M., Watkins, O., Posada Fishman, S., Aljubeh, M., Thacker, P., Fauconnet, L., Kim, N. S., Chao, P., Miserendino, S., Chabot, G., Li, D., Sharman, M., Barr, A., Glaese, A., & Tworek, J. (2025). GDPval: Evaluating AI model performance on real world economically valuable tasks. arXiv. https://arxiv.org/abs/2510.04374