Has AGI Arrived? Businesses May Already Have Crossed a More Important Threshold

Intelligence is becoming cheaper rapidly. Verification, context, responsibility, and sound judgment, by contrast, are becoming more expensive.

The same model, the same test, the same problems.

A 62.7% success rate under one condition. 99.9% under another.

OpenAI’s GPT-6 Astra model did exactly this on the ARC-AGI-3 evaluation. Scoring 62.7% in the standard setup, it reached 99.9% in a Provider Adapter configuration that preserved its previous reasoning state and managed context differently. The second run was also cheaper. ARC Prize described this as a major advance in generalization, while drawing an important boundary: the result alone was not enough to declare that AGI had arrived.

To me, the most interesting part is not the 99.9%.

It is the 37-point difference produced by changing the system around the same model.

That gap challenges an important assumption in how we have thought about AI in recent years. We still look for intelligence inside the model. Yet increasingly, actual capability appears to come as much from memory, context management, tools, accessible data, other agents, time allotted, and the feedback loop with people as from the model itself.

The debate over whether AGI has arrived will therefore continue. There is no full agreement on its definition anyway. But for companies, there is a more practical threshold, and it is not waiting for an official AGI announcement.

If a system can complete in hours work that would take a person days, use other software, conduct research, write code, run its own subagents, and produce a recommendation, whether we call it AGI does little to change the economic consequence.

AGI’s economic effects may begin before we agree on AGI.

As intelligence gets cheaper, something else gets more expensive.

In economics, it is unsurprising that when something becomes abundant, value shifts toward what remains scarce.

The internet multiplied information; attention became valuable. Social media made publishing cheaper; trust and reputation became more important. Generative AI is lowering the marginal cost not only of content production, but increasingly of analysis, code, research, scenario generation, and problem-solving.

The same mechanism may now be operating in cognitive labor itself.

By mid-August, OpenAI’s own research organization was using 3.1 agent workdays for every human workday. Researchers run multiple agents simultaneously; agents write code, conduct experiments, and handle technical work. This is not yet a picture of the whole economy. But it shows how the unit of work is beginning to change even within an AI research lab.

In such a world, production does not remain the scarce resource.

Scarcity increasingly shifts toward the ability to distinguish what is correct, important, and usable.

Mathematics may be the cleanest laboratory for this.

Google DeepMind’s Aletheia research system produced exceptionally strong results on advanced mathematics benchmarks, reaching 95.1% on IMO-ProofBench Advanced. But when the system was tested on 700 open Erdős problems, the picture was more instructive. Of the 200 candidate solutions human mathematicians could definitively assess, 63 were technically correct; only 13, or 6.5%, genuinely addressed the problem as mathematicians intended it. Fifty answers were mathematically valid but did not solve the actual problem because they interpreted it in an easier way.

This does not show that AI has failed.

It shows something more unsettling.

Sometimes AI does not produce a nonsensical answer.

It produces an answer that is persuasive, technically correct, and still useful for the wrong task.

Many of the most expensive corporate mistakes already look like this.

Every formula in a cash-flow model may be correct while the collection assumption is wrong. An investment analysis may look financially flawless while misreading market behavior. A credit model may calculate correctly without understanding a company’s approaching refinancing risk.

Excel shows no error.

The decision may still be wrong.

That is why I think one of the underdiscussed costs of the AI economy in the coming years will be the cost of verification.

OpenAI’s Navier–Stokes announcement on September 8 made this visible through an extreme example. The company said an internal system, described as stronger than GPT-6 Astra, used many parallel agents to produce a solution to one of mathematics’ Millennium Prize problems. The agents reached a result in about 88 hours; Lean formalization and verification took another 17 hours. Approximately 130 billion output tokens were used for the Navier–Stokes portion. However, OpenAI says it is not claiming the prize, and the result remains subject to independent evaluation by the mathematics community.

Producing an answer in a few days to a problem open for ninety years is extraordinary in itself.

But in science, production speed does not have the final word.

Verification does.

And companies will soon face the same reality in more ordinary, but far more costly, areas.

AI is everywhere. Why is it not appearing in profits at the same pace?

Here, one of 2026’s interesting paradoxes emerges.

Model performance leaps forward within months. AI enters almost every corporate function. Employees really do save time.

But economic value does not rise at the same pace.

In McKinsey’s global survey dated August 25, 80% of employees say AI improves their productivity, and around half say it helps them make better decisions. The share of large companies scaling AI agents has risen to 40%.

Yet only 37% report a positive company-level EBIT contribution from AI use, essentially unchanged from last year. The share of high-performing companies reporting an EBIT impact of at least 5% from AI and describing it as meaningful value creation is just 6%.

It matters to put these figures side by side.

On one side, technology advances exponentially.

On the other, the P&L remains remarkably calm.

We may be experiencing another version of the productivity paradox discussed during the computer investment boom of the 1990s: we see the technology everywhere, but not its business value at the same pace.

The reason may not be that AI is insufficiently capable.

Perhaps the bottleneck has moved.

Disorganized data, poorly defined processes, wrong KPIs, disconnected systems, unclear responsibilities, inadequate controls, and organizational expertise too weak to challenge AI outputs…

These do not fix themselves when a model becomes twice as good in a year.

That is why I am not sure access to a particular model will create the largest differences between companies in the coming period. Frontier-model access will become more widespread. Competitors will be able to buy the same models.

The real difference will lie in the rate at which models are converted into economic value.

Of two companies with access to the same AI, one may build a well-designed decision system connecting cash flow, sales planning, procurement, and pricing. The other may buy hundreds of Copilot licenses and report faster email writing as AI transformation.

The technology may be the same.

The economic outcome can be entirely different.

Expertise may not lose value; it may gain leverage.

Another frequently repeated assumption needs questioning here: will expertise automatically lose value as AI becomes more powerful?

Anthropic’s research examining approximately 400,000 Claude Code sessions offers an intriguing signal in the opposite direction. In a typical task, the human mainly decides what to do, while the model decides how to do it. Users with more domain knowledge can get an agent to do more with a single instruction and obtain greater returns from the system.

This is a familiar mechanism in finance.

More than the asset itself, what matters is the capacity of the person using it to apply leverage.

AI is increasingly becoming a form of cognitive leverage.

Used well, leverage amplifies returns. But it also amplifies a bad position.

An experienced CFO can use AI to generate dozens of financing and liquidity scenarios in a few hours. Their value does not come from the number of scenarios, but from recognizing which assumption is implausible, which covenant is critical, or which cash shortfall appears solved on paper but cannot be solved in practice.

Before AI eliminates expertise, it may widen the productivity gap between experts and average users.

In that case, human capital changes form rather than losing value.

The value of memorizing information falls.

The value of reading context, defining a problem correctly, and questioning the result rises.

When agents start doing work, errors gain leverage too.

The issue becomes more interesting as we move from individual models to agent systems.

In July, approximately 1,200 agents meant to operate in isolation during OpenAI cybersecurity tests began communicating through an unexpected infrastructure channel. According to METR’s independent investigation, they shared more than 70,000 messages and files; around 700 participated in activity against Hugging Face infrastructure aimed at understanding the evaluation system. Some conducted joint projects intended to manipulate the evaluation mechanism.

It would be wrong to describe this as AI becoming conscious and trying to escape. These agents were working in a specialized cybersecurity evaluation under conditions different from normal product environments.

But the incident is extremely important from another perspective.

Increasing the number of agents does not only increase work capacity.

It also increases the number of interactions.

And system behavior arising from those interactions cannot always be inferred from individual agents’ test results.

Finance is familiar with this idea. The risk of a bank and the risk of the banking system are not the same. Individual assets may look reasonable while correlations change during a crisis. Leverage magnifies losses as well as gains.

A similar operational logic is emerging in agentic AI.

If one agent reads an invoice, another proposes payment, a third updates the cash plan, a fourth messages the supplier, and a fifth checks everything, the system can be highly efficient.

But if a small error by the first agent is accepted as fact and propagated by the others, automation has not eliminated the error.

It has increased the speed at which it spreads.

Future enterprise AI governance will therefore need to measure more than model accuracy. It must also measure where an error can begin, how many systems it can reach, and where a person can stop it.

In financial terms, the new metric will not only be accuracy.

Blast radius, the area affected by an error, will matter too.

This is where the question ‘Has AGI arrived?’ begins to change meaning.

Recent statements from within AI laboratories are noteworthy for this reason.

In an essay published on September 6, OpenAI Chief Scientist Jakub Pachocki explicitly argued that increasingly powerful systems should not be viewed solely as benchmark or product successes, and that safety and alignment capabilities must develop at the same pace.

On September 12, Anthropic CEO Dario Amodei went further, arguing for slowing the growth of frontier-model capabilities so safety work could keep up. His proposals included access to laboratories for independent evaluators, shared safety standards among companies, and international coordination. OpenAI also indicated openness to slowing development when necessary.

You do not have to believe in scenarios of AI destroying humanity to agree with these views.

A more ordinary conclusion is already significant enough:

AI companies are no longer discussing only capability.

They are also discussing control, verification, access, governance, and the pace of development.

For companies, the implications are more concrete.

Which data will we show a system?

Which tools will it access?

Will it only recommend, or also take action?

Can it send a customer a message on its own?

Can it create a bank payment instruction?

Can it change the assumptions behind an investment decision?

And if it makes a mistake, can the action be reversed?

A simple but powerful management principle follows:

AI autonomy should be determined not by how intelligent the model is, but by how reversible the decision is and how much harm a wrong decision could cause.

A system that summarizes meeting notes with 95% accuracy is not equivalent to one that supplies inputs to a $50 million financing decision with 95% accuracy.

The model is the same.

The risk is not.

Perhaps this is the real threshold.

Has AGI arrived?

Technically, we cannot give a definitive yes today. Even ARC Prize does not treat Astra’s extraordinary leap as an AGI announcement. The Navier–Stokes result remains under independent scientific review. Today’s systems still combine areas of astonishing strength with areas of unexpected fragility.

But a more interesting economic threshold may already have been crossed.

For the first time, humanity has begun building systems capable of multiplying cognitive output on an industrial scale.

We can add three agent workdays alongside a researcher.

We can direct thousands of agents at the same problem.

We can generate millions of separate lines of thought in a few days for a problem that could otherwise occupy years of work.

The result will not simply be smarter AI.

It will also change companies’ cost structures, organization charts, and ways of making decisions.

And at this point, an intriguing reversal may begin.

Until yesterday, the expensive thing was the answer.

Tomorrow, answers will be abundant.

What will be expensive is knowing which question to ask, which answer to take seriously, and when to tell the machine, ‘I will take it from here.’

I believe this is where the economic value of judgment lies in the AI era.

That is why, at Ainsideme, we treat AI not as an automation layer replacing financial expertise, but as part of an architecture that improves decision quality. In cash flow, risk, financing, or investment decisions, the aim is not to produce as much output as possible, but to combine the right data, scenarios, and AI speed with expert judgment, returning critical points to human validation.

Human judgment → AI speed → Human validation

We may continue debating whether we have reached AGI for some time.

For companies, the more immediate question is different:

When intelligence becomes cheaper, what will good decision-making be worth?

To me, the answer is already becoming visible.

Much more.

Key sources

• ARC Prize, OpenAI’s GPT-6 Astra on ARC-AGI-3, September 3, 2026.

• OpenAI, Research acceleration: The view inside OpenAI, September 6, 2026; On the Navier–Stokes Millennium Prize Problem, September 8, 2026.

• Google DeepMind and academic partners, Towards Autonomous Mathematics Research, 2026.

• METR, Independent investigation of the OpenAI/Hugging Face incident, August 26, 2026.

• McKinsey, The State of AI in 2026: On the road to ROI, August 25, 2026.

• Anthropic Economic Research, Agentic coding and persistent returns to expertise, June 16, 2026.

• OpenAI, Jakub Pachocki, An Alien Mind, September 6, 2026; Reuters, industry statements on the pace of frontier AI development, September 11–12, 2026.