OpenAI has never published a document called the AGI roadmap. There is no Gantt chart, no milestone calendar, no public checklist of capabilities waiting to be ticked off. What exists instead is a scatter of blog posts, conference remarks, hiring pages, investor decks and offhand podcast lines that, read together, sketch something roadmap-shaped.
Learning to read that scatter is most of the skill. The pieces are real, and some of them are specific: a five-level capability framework, a half-trillion-dollar compute buildout, a safety framework with named thresholds. Others are deliberately vague, and the vagueness is doing work.
The Five Levels, and Why They Aren’t a Schedule
In July 2024, Bloomberg reported that OpenAI had built an internal scale for tracking progress toward artificial general intelligence. It ran like this:
- Level 1 — Chatbots: conversational AI that answers questions.
- Level 2 — Reasoners: systems that can work through multi-step problems, comparable to a human with a doctorate who doesn’t use tools.
- Level 3 — Agents: systems that can take actions over extended periods on a user’s behalf.
- Level 4 — Innovators: AI capable of generating new ideas and discoveries.
- Level 5 — Organizations: AI that can run the work of an entire company.
It reads like a ladder. It isn’t one. The levels describe kinds of capability, not equal-sized steps, and the gaps between them are wildly uneven. Moving from Level 1 to Level 2 took roughly eighteen months from ChatGPT’s launch to the o1 reasoning models in September 2024. Moving from Level 2 to a genuinely reliable Level 3 — agents that don’t quietly break after forty minutes — has proved considerably harder than the framework’s tidy numbering suggests.
There’s a second problem. OpenAI’s public definition of AGI, still printed in its charter, is “a highly autonomous system that outperforms humans at most economically valuable work.” That is a threshold, not a ladder, and it’s slippery enough that OpenAI’s AGI definition is doing far more work than most people notice — governing everything from Microsoft’s profit share to when the board’s obligations change.
The Compute Layer Is the Actual Roadmap
If you want the closest thing to a dated commitment in OpenAI’s AGI roadmap, look at data centres rather than model capabilities. In January 2025, OpenAI announced Stargate, a joint venture with SoftBank, Oracle and MGX, with a headline figure of $500 billion in infrastructure investment over four years. The first site, in Abilene, Texas, is being built at gigawatt scale.
That is a roadmap in the only sense a company can be held to: land, power contracts, GPUs, construction timelines. Sam Altman said in December 2024 that OpenAI “knows how to build AGI traditionally” — meaning the scaling recipe, in his framing, is understood and the constraint is now industrial rather than scientific. Whether or not you believe that, the spending pattern tells you what OpenAI believes. Compute is the bet, and the buildout is the schedule.
It also explains why the company’s public dates have a habit of drifting sideways rather than forward. The full list of AGI dates OpenAI has given is worth reading end to end, because the pattern is less about being wrong and more about which dates were ever meant literally.
What the Roadmap Deliberately Leaves Out
Two things are conspicuously absent from anything OpenAI publishes about AGI progress: a safety-to-capability ratio, and an exit condition.
The Preparedness Framework, published in December 2023, does define severity levels for catastrophic risk — low, medium, high, critical — and commits the company not to deploy models above a certain threshold without safeguards. But it doesn’t say when that threshold is likely to be approached, or what happens to the roadmap if it is. The framework is a brake pedal with no published stopping distance.
The other omission is the transition plan. Altman has repeatedly described AGI as arriving gradually rather than as a single event — a “continued exponential,” in his phrasing, rather than a switch flipping. That framing is convenient. A slow arrival is one nobody can be blamed for missing, and it makes the whole question of readiness permanently deferred.
Three Signals That Actually Indicate Progress
Ignore demo videos. If you want to track the OpenAI AGI roadmap yourself, these are the measurements that mean something:
- Task horizon length. How many hours of continuous, uninterrupted useful work can an agent sustain before it needs human rescue? Doubling this is real progress; a slicker chat interface isn’t.
- Cost per completed task. A capability that works at $400 a task and fails half the time is a demo. At $4 and 95% reliability, it’s a product — and that shift is what makes Level 3 economically real.
- Error recovery. Can the system notice it has gone wrong and correct course? Most agent failures aren’t capability failures, they’re compounding-small-mistake failures.
None of these appear in press releases. All of them show up eventually in enterprise pricing and in the complaints of people who actually deploy the things.
How to Read the Next Announcement Without Getting Fooled
OpenAI ships updates constantly, and the language around them is engineered to sound like milestones. A model that “reasons” is not necessarily a Level 2 system by the internal definition. A “research preview” is not a product. A benchmark record is not a deployment.
Three questions cut through most of it. First: is this a capability or a product? Second: what’s the failure rate in the regime where it matters, not on the benchmark? Third: what did OpenAI explicitly not say? The silence around reading AGI roadmap updates carefully is usually more informative than the announcement itself.
Why the Dates Keep Moving
Every forecaster in this field has shifted their estimates at least once. Geoffrey Hinton, who left Google in 2023 to warn about the technology, has said AGI could arrive in five to twenty years — a range wide enough to cover most outcomes. Altman has floated “a few thousand days.” Demis Hassabis has talked about a decade. The spread isn’t because anyone is lying; it’s because the underlying question — how much more can scaling deliver — is genuinely unsettled.
The pattern across the industry is consistent enough to be a phenomenon in its own right. The AGI development timeline keeps slipping for structural reasons — benchmark saturation, data scarcity, the difference between capability and reliability — and those reasons don’t disappear because a CEO gives a new number at a conference.
The Question That Beats Every Roadmap
Here’s the thing about the OpenAI AGI roadmap: nobody outside the building has seen it, and the people inside have strong incentives to describe it in whichever way serves the moment — ambitious for investors, cautious for regulators, imminent for talent.
So don’t try to predict the date. Ask instead what would have to be true for the date to matter. A system that outperforms humans at most economically valuable work needs to be reliable enough that you’d hand it a task and walk away, cheap enough to run at scale, and contained enough that the failure modes are survivable. Those are engineering and governance problems, not marketing ones. They will be solved visibly or not at all, and the evidence will show up in pricing pages, incident reports and the hiring decisions of companies that tried it — long before it shows up in a blog post from OpenAI.
Watch those. They’re the real roadmap, and nobody can edit them retrospectively.