Ask ten people what OpenAI’s AGI research actually consists of and you’ll get ten answers, most of which are about a date. That framing misses the interesting part. The lab’s push toward artificial general intelligence is not a single project with a finish line; it’s four research programmes running at once, plus a construction business bolted on the side to feed them.
There are the big pretraining runs that everyone assumes are the whole story. There’s the reasoning line that moved the goalposts in late 2024. There’s the agent work, which is where the money is starting to show up. And there’s safety research, which has quietly changed shape twice in two years. Understanding how those pieces fit together tells you far more than any leaked launch date.
What OpenAI Means by AGI (and Why the Wording Matters)
OpenAI’s charter, drafted back in 2015, defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Every roadmap, every evaluation framework, every press release is measured against that sentence. Which is a problem, because the sentence is doing an enormous amount of ambiguous labour. Note that it says most, not all. Note that it says economically valuable, which quietly sets aside unpaid care work, parenting, and a lot of what people actually consider human. And note that autonomous is a sliding scale nobody has bothered to define.
If you want the full teardown, there’s a good piece on what those 14 words actually mean, because the definition is the load-bearing wall of the entire research programme. Change the definition and every safety threshold, every capability eval, and every promised milestone moves with it.
The Research Stack: Three Bets Running in Parallel
Bet one: scale still pays, but the returns curve is bending
GPT-4 was, by most external estimates, a mixture-of-experts system with roughly 1.8 trillion parameters and about 220 billion active per token. The follow-up generation got bigger and more expensive, so much so that its predecessor’s successor was pulled from general availability partly on cost grounds. Then the frontier model shipped in August 2025 arrived as a router, sending easy queries to a small fast model and only invoking the big one when the prompt looked hard.
That detail is more revealing than any benchmark. When the leading lab starts optimising routing instead of raw parameter count, it’s an admission that the old trick of simply training longer has hit a data wall. There is only so much high-quality text on the internet, and synthetic data is a patch, not a cure.
Bet two: thinking longer beats training bigger
The genuinely new axis arrived in September 2024 with the first reasoning model, trained with reinforcement learning to produce long internal chains of thought before answering. By December, its successor was scoring around 87% on the semi-private ARC-AGI set when given a large compute budget. Same underlying pretrained model, wildly different results, purely from letting it think for longer.
This is the most consequential shift in recent memory. Intelligence on demand means you can trade tokens for accuracy at inference time, which changes both the economics and the safety picture. A model that can spend an hour reasoning through a problem is a different kind of object than one that answers in a second, and it opens attack surfaces that simple input filtering never had to handle.
Bet three: agents are where research meets revenue
In July 2024, OpenAI published a five-level scale for tracking progress: Chatbots, Reasoners, Agents, Innovators, Organisations. At the time it read like marketing taxonomy. Eighteen months later, the product line maps onto it almost exactly. Browsing operators, deep research tools, and coding agents are all Level 3 in that rubric, and they’re the first AI products with a plausible argument for replacing hours of salaried work rather than assisting them.
Whether that’s AGI research or just good product work is a fair question. The honest answer is that at this stage they’ve become the same activity, because agentic tasks generate exactly the long-horizon, verifiable feedback that reinforcement learning needs.
Safety Research Grew Up and Then Fractured
Two years ago the safety story was clean: a dedicated superalignment team, founded in July 2023 with a promise of 20% of the company’s compute and an initial $10 million in resources, tasked with solving alignment for superhuman systems within four years. It was disbanded in May 2024, days after two senior researchers resigned, one of whom said publicly that safety culture and processes had taken a backseat to shiny products.
What replaced it is less centralised and arguably more operational. The Preparedness Framework, first published in late 2023 and revised since, scores models across tracked risk categories including biological threats, cybersecurity, persuasion, and model autonomy, and gates deployment on whether safeguards hold up. Third-party evaluators now get pre-release access. The lab has published alignment techniques that train models to reason explicitly about their own safety policy rather than pattern-matching to refusals.
Is that enough? Serious people disagree. The critique that lands hardest isn’t that the evaluations are badly designed, it’s that the same organisation deciding what counts as dangerous is the one that profits from deploying.
Compute Is No Longer Support Staff
You cannot separate OpenAI’s AGI research from its infrastructure deals. The Stargate project, announced in January 2025 with SoftBank, Oracle, and MGX, put a $500 billion figure on four years of data centre construction, starting with a site in Abilene, Texas. Later commitments pushed toward multi-gigawatt power purchases.
This isn’t corporate housekeeping. The research agenda is written by the compute budget. Which experiments get run, how many seeds, how much reinforcement learning on verified tasks, how many evaluation passes: all of it is a function of available FLOPs. When a company signs a ten-gigawatt power agreement, that is a research statement.
Reading the Timeline Without Getting Fooled
Executives have said AGI is a “few thousand days” away, that the company now knows how to build it as traditionally understood, and that superintelligence is a matter of a few years. Each of those statements is hedged by a definition doing quiet work in the background.
It’s worth tracking every AGI date OpenAI has given and what each is actually worth, because the pattern is consistent: internal targets shift, public framings soften, and reporters keep quoting the original number. There’s a reason for that drift, and it isn’t dishonesty so much as the structure of the problem. The assumptions behind a 2027 prediction and a 2100 prediction differ on almost nothing except how much you believe compute scaling converts into capability, and you can read the assumptions behind those two dates to see how little distance separates them.
Five Signals Worth More Than Any Date
If you want to track OpenAI’s AGI research seriously, watch these instead of the press tour:
- Agent revenue versus agent demos. Real deployment means real failure reports, and failure reports are the best capability data available.
- Inference price trends. Reasoning costs have fallen by more than an order of magnitude in under two years. Cheap thinking changes what gets built.
- Evaluation transparency. Third-party access, published thresholds, and disclosure of what happens when a model crosses them.
- Governance structure. The conversion to a public benefit corporation with foundation control and a large Microsoft stake reallocated real authority. Read the fine print on what the safety commitments now require.
- Compute contracts signed versus megawatts delivered. Announcements are cheap; energised racks are not.
For a broader read on which parts of the published plan are engineering and which are positioning, the AGI roadmap breakdown is a useful reality check on the milestones themselves.
The Unresolved Problem Is Deployment, Not Capability
Here’s what the research doesn’t yet answer. A system that can outperform humans at most economically valuable work, once it exists, has to be integrated into an economy that runs on liability, accountability, and someone being responsible when things go wrong. No scaling law solves that. No benchmark measures it.
The uncomfortable possibility is that the capability arrives before anyone has designed the institutions to hold it. Companies are currently writing their own risk thresholds, auditing themselves, and publishing summaries they choose the shape of. That arrangement worked fine when the models were autocomplete. It gets harder to defend when the models are booking flights, filing tickets, and writing production code without supervision.
The most useful thing you can do as an observer is stop asking when AGI arrives and start asking what would have to be true for the answer to matter. Which evaluations would have to be public. Which decisions would have to sit outside the lab. Which capabilities would have to be demonstrated before deployment rather than after. Those are answerable questions. On the current evidence, nobody with the authority to answer them has committed to much.