The AI Race Shifts to Cost and Action. Is AGI Any Closer?

Three launches on Nov. 9 point the same way. Anthropic cut the price of its top model by 20%. Meta released a model that costs $0.10 per million input tokens. NVIDIA showed a robot model that learns from human video. None of these is mainly about a bigger brain. They are about making AI cheaper and putting it to work.

That fits a larger question. Is artificial general intelligence, or AGI, close? Lab leaders say it could arrive within a few years. Many academic researchers say today’s approach will not get there. This piece sets out both sides and what each would need to prove.

The race is changing shape

For years, the main contest was who could build the most capable model. Today’s news suggests a second contest is growing: who can deliver useful work at the lowest cost, including work in the physical world.

Price is part of this. Fable 5.2 costs $8 per million input tokens, down from $10. Ember costs $0.10. When an agent run costs cents, people try it on thousands of tasks, and that produces the evidence, failures included, that settles arguments.

Three launches in one day is thin evidence, so read this as our interpretation, not a measured trend. It still matters for the AGI debate. If cheap agents, meaning AI programs that take actions on their own, do more real work, we will learn their limits from daily use and not only from benchmarks.

First, what do we mean by AGI?

No one agrees. Some definitions mean a system that can do any task a typical person can do. Others mean one that does every task better than any human. A RAND report on AGI forecasts warns that AGI is not a single threshold. It says the wording of a question alone can move expert answers by decades. Prediction markets often use lower bars than these definitions, so their dates run earlier.

Think of asking when a car will be “fast enough.” Fast enough for a city street, or for a race track? The answer changes with the definition, long before the engine changes.

The case for “soon”

Dario Amodei, CEO of Anthropic, said at a January 2026 forum session that models could match Nobel-level work across many fields by 2026 or 2027. His reason is a loop: models that write code and do AI research help build the next models faster.

Demis Hassabis, CEO of Google DeepMind, is a little more cautious. In January he gave a 50% chance of human-level cognitive skills by the end of the decade. In May he told Axios that his central estimate is still 2030, with 2029 now possible. He said agents work in practice and that leading labs look to be on the right technical path.

A third signal is METR’s measure of how long a software task an AI agent can finish alone. RAND notes that it has doubled about every seven months, possibly faster. It covers software tasks only, so it does not describe work in general.

Anthropic’s launch claims fit this story. The company says Claude now writes over 90% of merged changes in its core codebase, with human guidance. That is a company claim, not an independent finding.

The case for “not this way”

The skeptics are not a fringe. In a 2025 survey tied to AAAI, a leading society of AI researchers, 76% of 475 respondents said that scaling up today’s approaches is unlikely or very unlikely to produce AGI. Only 23% said they were focused on pursuing AGI directly.

Yann LeCun, a Turing Award winner, argues that language models cannot predict the results of actions, which agents need. He says AI needs “world models,” systems that learn how the physical world works. Gary Marcus has made a similar case for years. RAND groups such views as “paradigm shift” thinking: deep learning is necessary but not enough, and AGI would come in the 2030s, the 2040s, or later.

Some early optimists have also slowed down. The AI Futures Project, which wrote the well-known “AI 2027” scenario, said at the end of 2025 that things were going “somewhat slower” than expected. Its lead author now puts his estimate at about 2030, with a lot of uncertainty.

A third group, which RAND calls economic skeptics, accepts that the technology may improve fast. They expect the economy to change slowly, because of deployment bottlenecks, adjustment costs, and politics. For them, even a very capable system would take years to show up in jobs and output.

What the surveys say

The largest survey of AI researchers, run in 2023 with 2,778 respondents, gave a 50% chance that machines beat humans at all tasks by 2047. That was 13 years earlier than the same group’s 2022 answer. Metaculus, a public forecasting site, moved from a median of about 2070 in 2020 to about 2033 in early 2026, according to RAND.

So timelines keep shrinking. But RAND lists real weaknesses. Surveys draw more optimists. Experts have no track record on AGI. The firms spending the most are also the ones making the boldest claims. RAND’s summary is that the direction of change is solid, while exact dates are not.

Where both sides agree

Almost everyone says today’s systems are not AGI yet. Hassabis names a missing skill: coming up with new questions and hypotheses, not just answering them. He also said it is unclear whether the self-improvement loop can close without people in it.

He made another point that is easy to miss. Code and math are easier to automate because answers can be checked. Natural science needs slow experiments to check a result. That gap alone can stretch a timeline.

So the real disagreement is about the size of the gap. One side sees one or two missing ingredients. The other sees the need for a new paradigm.

What to watch next

Arguing over dates is a weak use of time. RAND points to signs that would force people to change their minds. We would watch three:

1. Task length. Do agents keep extending how long a task they can finish alone?
2. Real projects. Can they complete software projects that take a human team weeks?
3. AI building AI. Does the share of AI research done by AI systems pass half?

We would add one more: whether cost cuts like today’s let many more people run agents at scale. Real use will test the claims faster than any forecast.

Our view is simple. The honest answer today is “we do not know,” and the definition matters more than the date. A reader can plan for a range of outcomes without picking a side.

Disclosure: Ad Astra’s drafts are written with the help of Claude, which Anthropic makes. This piece quotes Anthropic’s CEO and company claims, which we mark as such.