In March, the ARC Prize Foundation released a test where untrained humans solved 100% of the environments. The best frontier AI model on the planet scored below 1%.
A month before that, Sam Altman stood on a stage in New Delhi and said early versions of superintelligence might be a couple of years away.
Both of those are true right now. That’s the reason the AGI levels framework still earns its place. It’s the only public map that lets you hold “this is moving insanely fast” and “this is nowhere near finished” in the same hand.
I wrote about the AGI levels when Level 2 first landed. Two years on, the picture looks different.
What Are the AGI Levels?
OpenAI shared the five AGI levels with employees at an all-hands meeting in July 2024, and a company spokesperson confirmed the framework to Bloomberg two days later. The levels are Chatbots, Reasoners, Agents, Innovators, and Organizations.
Each rung takes more human direction out of the loop. Level 1 waits for your prompt. Level 5 runs the company.
What makes the framework useful is that it turned a fuzzy argument into a ladder. Before this, OpenAI defined AGI as highly autonomous systems that outperform humans at most economically valuable work, which is a sentence you can argue about forever without resolving.
The ladder doesn’t fix the definition problem. It gives you rungs you can point at. When someone tells you AI is about to run your business, you can ask which level they mean, and the conversation gets specific fast.
The 5 AGI Levels, One Rung at a Time
Here are the 5 AGI levels, explained:
AGI Level 1: Chatbots
Level 1 is conversational AI. Systems that understand natural language and answer you in it.
ChatGPT, Claude, Gemini. This is the rung everybody already lives on, and it got solved so thoroughly that it stopped feeling like AI at all.
My original version of this post described Level 1 as where we are. That was accurate in 2024. It reads strange now.
AGI Level 2: Reasoners
Level 2 is problem-solving at roughly doctorate level, working from the model’s own reasoning rather than tools or lookups.
OpenAI’s o1 opened this rung in late 2024. Everything since has pushed it further. Altman noted in February that AI went from struggling with high school math to contributing to theoretical physics in just over a year.
Level 2 of the AGI levels is cleared. These models still hallucinate and still fail in odd places, and it’s cleared.
AGI Level 3: Agents
Level 3 is where AI stops answering and starts doing. Systems that take a goal, plan the steps, run them across your tools, and adapt when something breaks.
This is the rung the entire industry is fighting over, and it’s where the framework gets uncomfortable. The demos are extraordinary. The production deployments are a graveyard.
Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, driven by escalating costs, unclear business value, and inadequate risk controls. Look at what’s absent from that list. Model capability.
We’re on Level 3. We’re not through it.
AGI Level 4: Innovators
Level 4 is AI that produces knowledge that didn’t exist before. Original ideas. New science.
Drug discovery is the usual example, and there are real results there. AI systems propose candidate molecules and protein structures that human teams then validate.
Dario Amodei’s framing for this is a country of geniuses in a datacenter. He asks you to picture 50 million people arriving around 2027, every one of them more capable than any Nobel Prize winner. Bold call. He hedges it himself, and I’d hold it loosely too.
Level 4 has barely started. What exists today is AI speeding up human discovery, with a human still deciding what’s worth discovering.
AGI Level 5: Organizations
Level 5 is AI running an entire organization. Coordinating operations, making strategic calls, allocating budget, with nobody in the chair.
This is the top of the AGI levels, and Altman has said plainly that a superintelligence would eventually do a better job as CEO of a major company than any human executive, including him.
Nothing today is close. The nearest analogue is autonomous trading, which runs without human oversight inside a very narrow box with very clear rules. An organization is the opposite of a narrow box.
Which of the AGI Levels Are We On Right Now?
Level 1 is behind us. Level 2 is cleared. Level 3 is underway and going badly for most of the companies attempting it.
MIT looked at 300 enterprise AI deployments and found 95% delivered zero measurable return to the P&L.
Gartner expects 33% of enterprise software applications to include agentic AI by 2028, up from less than 1% in 2024, which tells you the tooling arrives whether or not the deployments work.
So the honest answer on the AGI levels is 2.5. We cleared reasoning. We’re mid-stride on agents with one foot on the rung and one foot in the air.
Levels 4 and 5 are theory.
The Benchmark That Says We Are Further Out Than It Sounds
Here’s the number I keep coming back to.
ARC-AGI-3 launched on March 24, 2026. It drops AI agents into game-like environments with no instructions, no stated goal, and no hints about what winning looks like. The agent has to explore, work out the objective, build a model of how that world behaves, and plan.
Untrained humans solve 100% of them.
Zero model names left in the file. The line now reads:
Every frontier model tested, from OpenAI, Google, Anthropic and xAI, scored below 1%. The best of them did not clear half a percent.
Those same models scored in the 60s and 80s on the previous version of the test. The benchmark deliberately avoids language and external knowledge, so the collapse is about adapting to something genuinely unfamiliar.
There’s a $700,000 grand prize for the first agent that scores 100%. Submissions close in November, so it is still sitting there.
Read that against the AGI levels, and Level 3 gets a lot more honest. An agent that can’t work out an unfamiliar environment on its own is an agent that needs a human building scaffolding around it every single time.
Nobody Agrees on Where the Finish Line Is
The most telling development in this whole story had nothing to do with a model release.
In October 2025, Microsoft and OpenAI rewrote their partnership. The old contract said that if OpenAI declared AGI, Microsoft’s rights to the technology would evaporate, and OpenAI’s board could pull that trigger alone.
The new agreement states that any AGI declaration by OpenAI now gets verified by an independent expert panel.
Sit with that. Two of the largest companies on earth looked at the word AGI, decided it was too slippery to leave in a contract as a self-declared event, and hired referees.
The definitions keep sliding underneath everyone. Reporting suggested the practical trigger buried in that contract was financial: $100 billion in profits. Altman has called AGI a very sloppy term. Demis Hassabis, speaking at the same summit, put the change at 10 times the Industrial Revolution, unfolding over a decade rather than a century. Yann LeCun says current architectures will never get there at all.
The AGI levels are a map drawn by one company, toward a destination that same company keeps redrawing. Useful. Not gospel.
What the AGI Levels Mean for Your Business
Here’s the practical read.
Level 2 is sitting on your desk right now, cheap, and most businesses are still using it like Level 1. Typing questions into a chat box when they could be handing over analysis, planning, and decision support.
Level 3 is where the money is and where the failures pile up. That 40% cancellation forecast traces back to governance, cost, and undefined value. Companies point an agent at a broken process and hope the agent sorts it out. It doesn’t.
That’s the lesson buried in this ladder. Every rung assumes the rung below it is solid. An agent running on top of a process nobody documented fails in ways that look like a model problem and are really a plumbing problem.
I lived this one. Express Writers had close to 100 people and 7 years of process debt. When I built First Movers with 2 people, I built the systems first and the automation second, and we hit the same revenue milestone in under a year.
Same lesson, different scale.
Getting Ready for the Next Rung
You don’t need to predict when Level 5 lands. You need to be honest about which rung your business is standing on.
If you’re still copy-pasting into a chat window, Level 2 alone gives you back hours you don’t know you’re losing. Thaddeus Tondu at On Purpose Media recovered 250+ hours a month using tools that already exist.
If you’re reaching for agents, the work is unglamorous, and it’s the whole game. Map the process. Clean the data. Define what the agent can touch. Decide who’s accountable when it goes sideways.
That’s what we build at First Movers. If you want it done for you, book a call with our team. If you’d rather build it yourself, AI Labs has the 40+ courses and the community to get you there.
The AGI levels move whether or not you’re ready. Standing on Level 2 while everyone argues about Level 5 is the cheapest advantage on the table.
AGI Levels FAQ
What Are the 5 AGI Levels?
The 5 AGI levels are Chatbots, Reasoners, Agents, Innovators, and Organizations. OpenAI shared the framework internally in July 2024 to track progress toward artificial general intelligence, with each level removing more human direction than the one below it. Level 5 is where the ladder ties to AGI itself.
What AGI Level Are We At in 2026?
Roughly 2.5. Level 2 reasoning is cleared, with frontier models now doing research-level mathematics. Level 3 agents exist in production at a small fraction of companies, and Gartner expects more than 40% of agentic projects to be canceled by the end of 2027. Levels 4 and 5 remain theoretical.
Has AI Reached Level 3 on the AGI Levels?
Partially. Agentic systems ship inside major enterprise platforms and handle real multi-step work. They also depend heavily on human-built scaffolding, which is why ARC-AGI-3 results show frontier models scoring under 1% on novel environments that humans solve every time.
When Will AI Reach Level 5 of the AGI Levels?
Nobody knows, and the people closest to it disagree loudly. Altman suggests early superintelligence within a couple of years. Hassabis says AGI is maybe five years out. LeCun argues current architectures never get there. Treat any specific date as a bet, not a forecast.
Are the AGI Levels an Official Roadmap?
They’re OpenAI’s internal framework, first reported by Bloomberg rather than announced. That makes them a useful shared vocabulary and a company’s own map of its own progress. Other labs use different taxonomies, and DeepMind published a competing five-tier model of its own.
