When the frontier learns, what remains proprietary?

The strategic AI question is not simply whether a model can perform a task. It is how much output depends on knowledge, how much of that knowledge is already common, and whether operating the business creates proprietary know-how faster than the frontier makes equivalent know-how broadly available.

Astra makes the question immediate

On September 3, 2026, OpenAI released GPT-6 Astra, describing it as state of the art across computer use, software engineering, cybersecurity, science and professional work. The release matters here not because any single benchmark settles the economics, but because it is a fresh shock to the amount of high-value work that can be performed from the common frontier.

Latent Space's early-access testing is even more pointed. After reporting more than 20 billion tokens of testing, the authors conclude that OpenAI has trained a model “capable of automating much of their own AI Engineering.”

In the framework below, Astra is best understood as a potential increase in both H — effective common capability — and λ — the rate at which formerly scarce know-how becomes common. It does not prove that either parameter is large in every workflow. It makes the empirical question much harder to ignore.

Start with Romer: knowledge is a productive input

Traditional growth models treat technological progress largely as an external productivity term. Paul Romer's 1986 paper, “Increasing Returns and Long-Run Growth”, makes knowledge central to the growth process: private activity can create knowledge, and knowledge can generate increasing returns and spillovers.

For the enterprise, the useful extension is to separate the knowledge that is broadly available from the knowledge that remains economically specific to the firm.

Each exponent measures the percentage change in output associated with a 1% increase in its input, holding the others constant. Thus α captures output's sensitivity to capital and ν captures its sensitivity to labor, while β and γ capture sensitivity to common and proprietary knowledge respectively.

The exponents are not assumed to sum to one. That is deliberate. They are empirical questions about the production economics of a company, sector or workflow. A high β means common knowledge strongly affects output. A high γ means firm-specific knowledge remains economically important. If both are small, knowledge may not be the binding constraint at all.

Frontier AI changes the boundary between common and proprietary

Common and proprietary knowledge should not be double counted. Once know-how becomes generally available, it belongs economically in H, not in the residual proprietary stock R.

The firm therefore faces two competing flows: operating experience creates new proprietary know-how, while the frontier makes some existing know-how common.

Frontier AI can increase λ by making dispersed knowledge cheaper to retrieve, synthesize, reason over and execute. The result is economically ambiguous: the firm becomes more productive because H improves, while some of the scarcity value of R disappears.

A proprietary learning system creates a growing differentiated knowledge stock only when the firm can generate useful new know-how faster than the frontier commoditizes equivalent know-how.

Different sectors should have different answers

This framework implies there is no universal enterprise AI strategy. The answer depends on the elasticities.

Coding: a plausible high-β case

Coding may be close to a canonical high-β activity. The common knowledge stock is unusually deep: programming languages, libraries, APIs, algorithms, documentation, design patterns and open-source repositories. Outputs are also unusually machine-verifiable through compilation, tests and runtime behavior.

In a controlled GitHub experiment, developers with Copilot completed a standardized coding task materially faster than those without it. The important economic point is not the exact percentage; it is that a broadly available model could deliver a large productivity gain without requiring years of firm-specific learning. GitHub reports a 55% faster completion rate in that experiment.

Astra pushes the example further. Latent Space reports from its early-access testing that the model can choose and train models, label data, instrument and read logs, deploy and debug systems, manage subagents and run evaluation workflows — tasks that sit inside the specialized occupation of AI engineering itself. Their conclusion is that Astra can automate much of OpenAI's own AI-engineering work.

That does not make proprietary software knowledge irrelevant. Internal architecture, codebases, business logic and production history can still make γ meaningful. But the Astra release is evidence that β can be very large even in technically sophisticated work, while λ can rise quickly enough to compress scarcity rents in occupations that looked highly defensible only months earlier.

Customer support: β and γ may both matter

Technical customer support has a different structure. General language and reasoning capability matters, but output also depends on product-specific behavior, customer configurations, historical failure modes, escalation patterns and which interventions actually resolved earlier cases.

Brynjolfsson, Li and Raymond studied a generative-AI assistant across 5,179 customer-support agents. They report a 14% average productivity increase and a 34% increase for novice and lower-skilled workers, with suggestive evidence that the system disseminated practices associated with stronger workers. That is endogenous learning in recognizable form: localized expertise becomes reusable organizational capability.

Two sides of the same knowledge market

Mercor's AI-agent environment work illustrates one side of the mechanism. Experts recreate realistic work, define judgment calls, write reference answers and specify criteria that separate strong responses from plausible-but-wrong ones. That converts expert know-how into reusable evaluation and training assets for AI systems.

Trajectory describes the complementary enterprise move: capture production traces, corrections and outcomes; derive evals from real traffic; replay tasks; and make agents better with use.

What enterprises should measure

The theory becomes operational when the variables become measurable at the workflow level:

• β: how much performance improves simply by substituting a better generally available model.
• γ: how much additional performance depends on firm-specific knowledge after controlling for the frontier.
• φ(E): how efficiently operating experience becomes reusable capability.
• λ: how quickly today's differentiated know-how becomes tomorrow's common model behavior.

Those parameters should vary sharply across workflows. Estimating them is a more rigorous way to decide where enterprises should build proprietary intelligence, where they should simply rent the frontier, and where knowledge is not important enough to justify either.

Related Proverify thinking