Ask five people at OpenAI what artificial general intelligence actually is and you will get five slightly different answers. Ask them when, and the spread stretches from “a few years” to “decades.” That gap between definition and date is the most honest thing about the whole debate, and it is the right place to start any serious look at OpenAI’s path to AGI.
The company’s charter describes the goal as “a highly autonomous system that outperforms humans at most economically valuable work.” Fourteen words, and every one of them is doing heavy lifting. Most economically valuable work. Outperforms by what measure. The definition stays loose on purpose, because a tight one would be a liability the moment a lawyer or a regulator read it closely.
The five-level ladder OpenAI actually uses
In July 2024, Bloomberg reported that OpenAI had been tracking its own progress against a five-tier internal scale. It has since become the closest thing the company has to a published roadmap, even though it was never released as one.
- Level 1, Chatbots: conversational systems that answer but do not act. GPT-3.5 and the first GPT-4 lived here.
- Level 2, Reasoners: models that think before answering, with human-level problem solving inside narrow domains.
- Level 3, Agents: systems that take multi-step actions on a user’s behalf over hours or days.
- Level 4, Innovators: systems that generate genuinely new ideas rather than recombining what they have already seen.
- Level 5, Organisations: an AI capable of running the work of an entire company.
Read that list quickly and it looks like a plan. It is not one. No dates are attached, no threshold defines what crossing a line means, and nothing says the levels are sequential. OpenAI researchers have described the categories in interviews as a communication aid rather than an engineering specification.
Level two turned out to be a different kind of scaling
The o1 release in September 2024 showed where the company thought the next gains were hiding. Instead of a larger model trained on more text, o1 spent more compute at the moment of answering, generating a long internal chain of reasoning before writing a single word of output.
o3, previewed in December 2024, pushed that idea hard. On ARC-AGI, a benchmark built to resist memorisation, it scored around 87 percent at high compute, above the roughly 85 percent human baseline, at a reported cost near $20 per task. On FrontierMath, a set of research-level problems, it moved from single digits to roughly a quarter correct.
Why that flips the economics
For a decade the working assumption was that intelligence gets cheaper every year. Reasoning models broke it. Better answers now cost more, because you pay for thinking time as well as for parameters. That single change reshaped product design, pricing pages and the entire business case for building on top of a frontier model.
The pretraining question nobody has fully answered
GPT-4 arrived in March 2023 and set a bar competitors needed a year to reach. What followed looked more like refinement than revolution. GPT-4o made models faster and multimodal. GPT-5 in August 2025 behaved as a routing system, deciding when to think hard and when to answer instantly. Useful engineering, but not the same kind of leap.
The constraint is data. High-quality human text is finite, and the well is close to dry. Synthetic data helps, and OpenAI leans on it heavily, though training on model output risks a slow drift toward blandness. The live question is not whether scaling laws still hold. It is whether the next unit of capability comes from more pretraining, more test-time compute, or better reinforcement learning environments. Each points to a very different company. If you want to judge how much has genuinely changed between generations, the real differences between each new wave of AGI models are worth examining closely.
Agents and the reliability tax
Operator, launched in January 2025, could click through websites and complete bookings. Deep Research, weeks later, could spend twenty minutes reading sources and return a cited report. Both are genuine level-three behaviour. Both also fail in ways that are trivially easy to reproduce.
The arithmetic is unforgiving. If an agent clears each step with 95 percent reliability, a twenty-step task succeeds about 36 percent of the time. Stretch it to fifty steps and success drops under 8 percent. Autonomy multiplies error, which explains why so many agent demos look magical and so few agent products survive contact with a real workflow.
Real progress will not look like a viral clip. It will look like task horizons getting longer. METR, the evaluation nonprofit, has measured how long a task frontier models can finish at 50 percent reliability and found the figure roughly doubling every seven months. Hold that rate and week-long autonomous projects land somewhere around 2027.
Roughly half a trillion dollars of conviction
None of this runs without power. Stargate, announced at the White House in January 2025 with SoftBank, Oracle and MGX, was framed as $500 billion of infrastructure across four years, starting with $100 billion. The first site rose in Abilene, Texas, on a plot measured in hundreds of acres and gigawatts of grid capacity.
That scale is the strongest argument for a short timeline, and also the biggest risk to it. Data centres take years to build. Grid connections take longer. An organisation committing that much concrete has effectively bet the company on models continuing to improve enough to justify the bill.
Safety, restructuring, and a definition that keeps moving
OpenAI’s governance has changed more than its models have. The superalignment team, formed in 2023 to work on controlling systems smarter than humans, was dissolved in May 2024 after Ilya Sutskever and Jan Leike left. A safety and security committee took over, the Preparedness Framework introduced risk tiers, and in 2025 some of its own thresholds were relaxed.
In October 2025 the corporate structure shifted again, with the for-profit arm becoming OpenAI Group PBC under a nonprofit foundation holding roughly 26 percent and appointing the board. The logic is capital, since research at this scale needs investors who expect a return. The cost is that the AGI clause in the Microsoft agreement, which governs when the technology is formally declared, now sits inside a far more complicated machine. Regulators are moving in parallel, and AGI policy has been shifting faster than the models it governs.
What would actually count as arriving
Here is the awkward part. The charter requires the board to declare AGI, and that declaration carries commercial consequences. In 2023, leaked terms suggested Microsoft’s access would end at a definition tied to $100 billion in profit. That is a financial trigger wearing scientific clothing.
Sam Altman has offered a looser benchmark repeatedly over the years: a system that does what a median human can do at most economically valuable work. It sounds concrete until someone tries to measure it. Which jobs. Which median. Does a model that writes better code than most developers but cannot reliably book a flight qualify.
Benchmarks will not settle it. Frontier models saturate exams within months of release, then stumble on tasks a ten-year-old manages without effort. Telling genuine capability gains from benchmark gaming is now a skill in its own right, and a large share of what gets called an AGI breakthrough is really a strong score on a narrow test.
Signposts that separate signal from noise
Stated timelines cluster hard. Some insiders point at 2027, others at a slow climb past 2040, and the assumptions behind the 2027 and 2100 estimates do almost all the work in that disagreement. Rather than pick a year and defend it, watch these instead:
- Task horizon. If the doubling trend holds, agents that run for weeks without human rescue are the clearest evidence that level three is closing.
- Cost per completed task. Not per token. A system that finishes a job for $5 beats a cheaper one that needs three attempts.
- Novel problems. FrontierMath and ARC-AGI matter precisely because they were not in the training set. Watch for new benchmarks that survive more than one release cycle.
- Whether OpenAI publishes its AGI criteria. A definition written down in advance would be worth more than any demo.
There is a useful discipline buried in all this. Most weeks, nothing happens. A benchmark tick, a pricing tweak, a paper. Every few months something shifts and the whole conversation resets. Learning to separate real laboratory progress from the noise around it is more practical than arguing about dates with strangers.
The ladder has five rungs, no schedule, and a definition that gets rewritten as the company climbs it. That is less evasion than honesty about an unsolved problem. The engineering is real. So is the distance between a polished demo and a Tuesday afternoon inside an actual business. If you want one number to follow, make it the task horizon, because that is where the gap between level two and level five genuinely gets closed.