The numbers first
In 2019, the median expert prediction for AGI arrival was around 2060. By 2022, after GPT-3 and the first wave of foundation models, it dropped to 2048. The 2023 wave — GPT-4, Claude 2, Gemini — pushed it to 2037. Today, September 2026, the Metaculus community median sits at 2031. The AI Impacts expert survey, which polls hundreds of researchers, lands in the same range.
Twenty-nine years of forecasted future, gone in seven. That's not experts getting excited about headlines. It's a response to specific, measurable things that happened.
What actually moved the needle
Three categories of evidence are driving this, not hype cycles.
Benchmark saturation. Tests that were designed to be AI-proof for decades are getting cleared years ahead of schedule. MMLU — a massive multitask language understanding test covering 57 subjects from law to physics — was expected to remain beyond reach until the mid-2030s. GPT-4 hit expert-level performance in 2023. GPQA, which tests PhD-level knowledge in biology, chemistry, and physics, was cleared by Claude 3 Opus in 2024. Each cleared benchmark is a calibration point. Experts update accordingly.
Coding and software engineering. SWE-bench — a benchmark of real-world GitHub issues requiring multi-file code changes — went from 2% solve rate in 2023 to over 50% by mid-2026. That's not a marginal improvement. Software engineering was the profession many researchers pointed to as the clearest proxy for general cognitive ability. Watching it collapse as a differentiator has been meaningful to people who set timelines.
Agentic task completion. This is newer and harder to benchmark cleanly, but the direction is clear. AI systems completing multi-day software projects, running research pipelines, autonomously iterating on experimental results — these are qualitative shifts that didn't exist at all in 2019. They compress the conceptual gap between "very capable language model" and "general-purpose cognitive system."
Worth noting: the forecasts are still spread across a wide range. Some credible researchers put AGI at 2027–28. Others at 2040+. The median is just where the probability mass sits — not a guarantee, and not a cliff edge. What's striking is that the entire distribution has shifted left, not just the optimist tail.
Why this isn't just herding
A reasonable objection: are experts just updating toward the AI hype cycle, following each other rather than the evidence?
Some probably are. But the researchers who've moved most dramatically — Hinton, Sutskever, the teams at Epoch AI — aren't doing it on vibes. They're pointing to specific capability discontinuities: the moment o1 showed that giving models time to reason before responding dramatically improved performance on hard problems. The moment coding benchmarks stopped being meaningful tests. The moment agentic systems started completing tasks that required sustained multi-step planning without human correction.
The forecasters who remain skeptical — Yann LeCun being the most prominent — aren't arguing that things haven't improved. They're arguing that the current architecture (transformer-based, trained on next-token prediction) has a ceiling, and that ceiling is lower than AGI. That's a legitimate position. But it's worth noting that the same architectural critique existed in 2020, and the benchmarks kept falling anyway.
What the score actually measures
The AGI proximity score on this tracker pulls from four weighted inputs: expert surveys (40%), prediction market consensus (25%), benchmark performance (25%), and lab roadmap signals (10%). The current score sits at 65 out of 100 — up from 62 when we launched in August.
The benchmark component is what's moved most in the last year. Expert surveys update slowly (major surveys run every 1–2 years). Prediction markets are reactive. Benchmarks are continuous, public, and hard to fake — and they've moved faster than the expert surveys have caught up with.
What comes next
The forecasts will keep moving as long as benchmarks keep falling. The ones to watch:
- ARC-AGI 2 — François Chollet's updated benchmark designed to resist the pattern-matching that cleared previous tests. If this falls quickly, it removes one of the strongest remaining "not yet" arguments.
- Autonomous scientific discovery — AI systems making genuine novel contributions to research, not just summarizing literature. AlphaFold was one instance. More are in progress.
- Sustained agentic performance — completing tasks over days or weeks without human correction, across domains. The infrastructure for this is being built now.
None of this means 2031 is certain. It means the evidence base for pushing the timeline out has gotten thinner with each passing year, and the researchers tracking it have noticed. For a profession-by-profession look at what this timeline means for your work, see the full job risk breakdown.
Frequently asked questions
How far away is AGI realistically?
The most credible current estimate puts the median around 2031 — that's where both the Metaculus community forecast and the AI Impacts expert survey land as of late 2026. "Realistically" is the key word: forecasters who dismiss the near-term timeline tend to be anchoring to pre-2022 intuitions that haven't updated against benchmark data. The researchers actively building these systems have moved their estimates in considerably faster than the academic consensus has.
How long until AGI is achieved?
At the 2031 median, that's roughly five years from now (as of 2026). But the more useful frame is that near-AGI capability effects — systems that can do most knowledge work autonomously — are likely to be felt two to three years before the formal threshold. That puts meaningful disruption arriving around 2028–2029 for people making decisions today.
How close is ChatGPT to AGI?
ChatGPT (GPT-4o and its successors) passes professional exams, exceeds median human performance on many reasoning benchmarks, and handles complex multi-step tasks — but it still fails on novel problem-solving tests designed to resist pattern matching, like ARC-AGI 2. The honest answer is: closer than most people realized two years ago, not there yet in the ways that matter for full autonomy. OpenAI's own internal probability estimate, reported in mid-2026, put their systems at roughly 70% of the way to their internal AGI definition.
How close are we to AGI in 2026?
The gap between current systems and AGI narrows along two dimensions: raw capability and autonomy. On raw capability, 2026 systems are doing things that would have been AGI-level predictions in 2019 — the goalposts moved. On autonomy (sustained, reliable, multi-day task completion without correction), the gap is larger. The 2026 picture is: capability is nearly there in many domains, but reliable autonomous agency across domains is not. That's where the remaining 2031 gap lives.
What comes after AGI — and what is ASI?
AGI (artificial general intelligence) refers to systems that match or exceed human-level performance across all cognitive domains. ASI (artificial superintelligence) is what comes next: systems that significantly exceed the best human performance — not just matching it. Most forecasters who model AGI arriving around 2031 expect the transition to ASI to happen faster than the transition to AGI did, because a general-intelligence system can accelerate its own development. For a deeper look at the AGI-to-ASI gap, see the AGI vs ASI breakdown.