AGI by 2027: who's predicting it and what would it actually take

The claim and what it actually means

In September 2026, we're inside the window Dario Amodei publicly flagged. He expected "powerful AI systems" to emerge in late 2026 or early 2027. That's not a casual comment — it appeared in Anthropic's regulatory filings and in multiple public interviews. The question is whether his definition of "powerful AI" is the same as what most people mean when they hear "AGI by 2027."

Spoiler: it isn't, quite. And understanding the gap between the claim and the common reading of it matters for how you plan the next 12 to 18 months of your career or your product roadmap.

The predictors and their actual claims

The key distinction: Amodei and the AI 2027 project define "AGI" or "powerful AI" around a very specific capability — AI that can automate AI research itself. That's the threshold that matters in their framework, because once AI is building better AI, the pace of development becomes self-sustaining. Jensen Huang and Greg Brockman are using the term more loosely, to describe AI that is transformative at scale, not a specific recursive threshold.

What AGI by 2027 would require

Under Anthropic's definition — which is the most precisely stated of the bunch — here's what a "late 2026 / early 2027" AGI would need to be able to do:

• Match or exceed Nobel Prize-level expertise across biology, mathematics, engineering, and programming — simultaneously
• Execute autonomous tasks over multi-day or multi-week timeframes with minimal human oversight
• Operate at 10x to 100x human cognitive speed
• Fully automate AI research and development — meaning it can design, run, and interpret experiments that improve AI systems without human involvement

GPT-6 Astra comes closest to the first two requirements. It scores at or above PhD-level on GPQA Diamond (96%), outperforms human experts in cybersecurity benchmarks, and can operate computers autonomously. But the third and fourth requirements — speed and self-directed AI R&D — remain undemonstrated publicly.

Ryan Greenblatt, a safety researcher at Anthropic itself, estimated the probability of this version of AGI arriving by early 2027 at roughly 6%. This from someone inside the organisation making the prediction.

"We expect powerful AI systems will emerge in late 2026 or early 2027 — systems that match or exceed Nobel Prize-level experts across most scientific disciplines."

— Dario Amodei, Anthropic CEO (regulatory filing, 2025)

The AI 2027 scenario: what it predicted, and how it's tracking

The AI 2027 project — built by a team including former OpenAI researchers — published a quarter-by-quarter roadmap to AGI in early 2025. It's worth examining because it's the most granular public prediction on record, and because we can now check parts of it against what actually happened.

The scorecard: AI 2027 was genuinely prescient on the qualitative dynamics — government friction, security risks, elite model access — and too aggressive on the quantitative timeline. That's not unusual for predictions that are mostly right: they compress the schedule.

The forecasters who think 2027 is too soon

For contrast, here's the current spread of credible forecasts and what they imply for a 2027 window:

The spread reflects genuine disagreement on two variables: how much compute matters relative to algorithmic progress, and whether current architectures can reach AGI or whether a new paradigm is required. The optimists (Amodei, Altman, Legg) are betting heavily on compute scaling and architectural improvements within the current paradigm. The skeptics (LeCun, Marcus, Gary Marcus's bet at 10-to-1 against specified AGI tasks by end 2027) believe current systems are missing something fundamental — grounded physical models of the world, continuous learning — that no amount of scale will provide.

What changes if AGI arrives by 2027

The most concrete consequence isn't "AI is generally intelligent." It's a specific cascade that the AI 2027 project maps clearly: AI that can do AI research → faster AI development → shorter gaps between capability generations. Under their scenario, the time between major capability jumps compresses from 18 months to 6 months or less, then to weeks, as AI-assisted research compounds.

Dario Amodei has stated this directly: AI will replace "all software developer work within a year" of such a system emerging. He doesn't mean every developer gets fired on day one — he means that the marginal cost of software development drops toward zero, which restructures the economics of every software business.

For builders and workers, the 2027 scenario means the window to adapt is very short. If the superhuman coder arrives by late 2027 and the AI research acceleration kicks in, you get roughly 12–18 months to position yourself on the right side of the capability curve before the next wave. That's not a comfortable runway — but it's a real one. See the one-person unicorn framework for one way to think about building into that window.

The honest assessment: very possible, not probable

The balance of independent forecasting evidence puts full AGI by 2027 in the 5–15% probability range. That's not negligible — a 1-in-10 chance of AGI within 15 months is genuinely alarming if you're not thinking about it — but it's well below what Amodei's confident framing implies.

The more likely outcome is a version of the AGI era that Brockman announced: not a hard threshold crossed by a specific date, but a progressive capability expansion that makes the question of "is AGI here?" increasingly semantic. The aggregate median forecast sits at 2031. That could move significantly based on what happens in the next 6–12 months of model development.

Whatever the exact date, the compression from a 2060 forecast to a 2031 median in less than a decade means one thing above all else: the people who planned as if AGI was a long-run problem are already behind. The people planning as if it's a short-run problem — even if they're slightly early — will be better positioned regardless of which exact year the threshold gets crossed.