How Many Levels of AI Are There? Every Ladder Compared

Here are three numbers about the same kind of system.

In November 2023, a Google DeepMind paper put ChatGPT at Level 1. In July 2024, Bloomberg reported that OpenAI had told its staff it was at Level 1 and close to Level 2. In May 2025, Anthropic launched Claude Opus 4 under ASL-3.

So is a 2025 chatbot level 1, level 2 or level 3? The answer is: none of those numbers can be compared. They come from three different ladders, and each one measures a different thing. DeepMind’s ladder measures how good a system is compared with skilled people. OpenAI’s reported ladder measures what kind of work it can do. Anthropic’s measures how dangerous it might be, and what protections it needs.

People talk about “level 3 AI” or “AGI” as if there were one staircase. There isn’t. I count at least six widely used ladders, measuring different things: generality, kind of mind, type of work, performance, autonomy and risk. This post takes each in turn: who made it, the levels, where it breaks, and where frontier AI sits on it in October 2026. (The deeper history of the word “AGI” is in What Does “AGI” Even Mean?.)

The six ladders at a glance

Keep one thing from that picture: the rows don’t line up. “Level 3” in one column has nothing to do with “level 3” in the next.

Ladder 1: Narrow, general, super (ANI, AGI, ASI)

Who and when. No single author; it grew from older ideas. In 1980, the philosopher John Searle, in “Minds, Brains, and Programs”, split “weak AI” (the computer as a tool for studying the mind) from “strong AI” (an “appropriately programmed computer really is a mind”). Searle’s split was about minds, not breadth, though “strong AI” later came to mean general AI loosely. “Artificial general intelligence” dates from 1997 (the story is in the AGI post). In 2014, Nick Bostrom’s book Superintelligence defined the top rung as “any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest.” The tidy three-letter trio of ANI, AGI and ASI was popularised for a general audience by Tim Urban’s Wait But Why essay “The AI Revolution” in January 2015.

What it measures. Breadth. How many different kinds of things can the system do?

Strengths. Everyone understands it in ten seconds, and the narrow/general distinction is real.

Criticisms. Three rungs is far too coarse. There’s nothing between “one task” and “human breadth”, which is where every interesting system of the last five years lives. It says nothing about how good a system is: superhuman AlphaFold and a mediocre general model share the bottom half. And “AGI” has no agreed test, so crossing the middle line is a matter of definition, not measurement.

Where AI sits in October 2026. Disputed. Frontier models are clearly not narrow: one model writes code, translates, does olympiad maths and reads images. Whether that’s AGI depends on the definition, and the people who build tests for it keep saying no. Nobody credible claims ASI, though it’s now in company names: Safe Superintelligence Inc. (June 2024) and Meta Superintelligence Labs (June 2025).

One newer attempt replaces the line with a score. “A Definition of AGI” (Hendrycks and colleagues, October 2025) scores models across ten areas of human cognitive ability and put GPT-4 at 27% and GPT-5 at 57%. I found no official score for newer models.

Ladder 2: The “four types of AI”

Who and when. Arend Hintze, then at Michigan State University, in a November 2016 article for The Conversation, “Understanding the four types of AI, from reactive robots to self-aware beings.” It’s now in countless explainers and course pages.

What it measures. The kind of inner machinery a system has: memory, a model of other minds, a model of itself.

Strengths. Memorable, and it points at real questions. Hintze’s own conclusion was modest: self-aware machines are far off, so focus on memory and learning from experience.

Criticisms. Popular but loose: a short magazine piece, not a standard, with no tests attached. Who decides whether a system “has” theory of mind? A 2024 Nature Human Behaviour study (Strachan and colleagues) gave theory-of-mind tests to GPT-4, Llama 2 and 1,907 people. GPT-4 matched or beat humans on several, such as false beliefs and indirect requests, but struggled to spot a social faux pas. The authors called this behaviour consistent with mental-state reasoning, not proof of it. Type IV, self-awareness, has no agreed test at all.

Where AI sits in October 2026. Comfortably Type II, with real but disputed signs of Type III on tests. No one has shown Type IV.

Ladder 3: OpenAI’s five levels (reported, not published)

Who and when. Bloomberg reported in July 2024 that OpenAI had shared this scale with employees at an all-hands meeting, as a work in progress that might change. According to the reports, leaders said the company was at Level 1 and “on the cusp” of Level 2.

This ladder is quoted everywhere as if it were official. As far as I could find, OpenAI has never published it as a formal framework. Sam Altman did use the framing in public, discussing o1 as the reasoning step and agents as the next at T-Mobile’s Capital Markets Day in September 2024. But the definitions come from press reports of an internal document.

What it measures. The kind of work a system can take on, from talking, to thinking, to acting, to inventing, to running a company.

Strengths. Concrete and product-shaped, and it matches the industry’s actual sequence: chat (2022), reasoning models (late 2024), agents (2025).

Criticisms. No published definitions or tests, from a company with a commercial stake in climbing it. The levels overlap, too: agents must reason, and “aid in invention” is a matter of degree.

Where AI sits in October 2026. Level 3 by most readings. Agents that take actions are real products: OpenAI launched ChatGPT agent in July 2025, and coding agents now work through long tasks on their own. Level 4 is contested. Models solve olympiad problems and help with research, but “innovator” is exactly the word whose meaning people argue about. I found no official OpenAI statement placing its 2026 models on this ladder.

Ladder 4: Google DeepMind’s Levels of AGI

Who and when. Meredith Ringel Morris and colleagues at Google DeepMind, including Shane Legg, published “Levels of AGI” in November 2023. It was presented at ICML 2024, and revised in September 2025.

What it measures. Each performance level comes in a narrow and a general version, so it’s a grid of performance by generality. A second ladder covers autonomy, from AI as a Tool to AI as an Agent. The AGI post covers the grid in detail.

Strengths. It separates how good from how broad, so AlphaFold (narrow, Level 5) and ChatGPT (general, Level 1) both have a sensible place. And it separates capability from autonomy.

Criticisms. Nobody has actually measured “percentile of skilled adults” across “a wide range of tasks”. And today’s systems are jagged, expert on some tasks and below a child on others, so one rung for a general system is hard to assign.

Where AI sits in October 2026. The paper’s 2023 placement was Level 1, “Emerging AGI”, for ChatGPT, Bard, Llama 2 and Gemini. The 2025 revision renamed a level but, as far as I can tell, re-placed no model. Frontier models now reach expert level on some tasks, like competition maths and coding, and did badly on others, like ARC-AGI-3’s novel games at launch. So: Level 1 general by the paper, arguably Level 2 on many tasks, no official re-rating. Disputed.

Ladder 5: Autonomy levels for AI agents

This one borrows from cars.

The model: SAE J3016. In 2014, SAE International published J3016, the standard behind the “Level 0 to Level 5” language for self-driving cars. Level 0 is no automation, 1 is driver assistance, 2 is partial automation, 3 is conditional automation, 4 is high automation within a defined area, and 5 is full automation in all conditions. It was last revised in 2021 (jointly with ISO). The useful idea is that a level belongs to a feature in a setting, not to a car: one vehicle can have a Level 2 highway assistant and a Level 0 parking sensor. Robotaxis like Waymo’s are commonly described as Level 4.

The AI version. In July 2025, Kevin Feng, David McDonald and Amy Zhang published “Levels of Autonomy for AI Agents” with the Knight First Amendment Institute at Columbia University. Their key move is to define each level by the role the human plays, not by how smart the agent is:

They also propose autonomy certificates: a third-party body would test an agent, given its model, prompts, tools and environment, and certify the highest level it may operate at, with re-certification when any of those change.

What it measures. How much control the human hands over. Crucially, the authors treat autonomy as a design decision, separate from capability. A very capable agent can be deliberately run at Level 1.

Strengths. It matches real deployments: the same coding agent can ask before every command or run on its own, and this ladder describes both. It gives a practical safety lever: limit the level, not just the model.

Criticisms. It’s new, and nobody issues autonomy certificates yet. A level is a setting, not a guarantee: an agent can still surprise you between check-ins. And unlike cars, no regulator requires anyone to declare one.

Where AI sits in October 2026. Everywhere, because it’s a dial. Frontier models can sustain long, unsupervised work: METR estimated an early Claude Mythos Preview, tested in March 2026, could handle tasks taking skilled humans at least 16 hours at 50% reliability, near the limit its tasks can measure. So capability supports Levels 4 and 5 for many software jobs. Whether an agent actually runs there is chosen by whoever deploys it, which makes this the most useful of the six for practical decisions. (For what makes something an agent at all, see What an “AI Agent” Actually Is.)

Ladder 6: Risk and safety levels

This family asks how much harm a system could enable, and what protections that requires.

Anthropic’s AI Safety Levels (ASL)

Who and when. Anthropic introduced ASLs in its Responsible Scaling Policy in September 2023, “modeled loosely” on the US government’s biosafety levels (BSL-1 to BSL-4) for handling dangerous biological materials. The idea: the more dangerous the capability, the stricter the security and deployment rules.

What happened since. On 22 May 2025, Anthropic activated its ASL-3 deployment and security standards alongside Claude Opus 4, the first time it had done so. It was explicitly a precaution: Anthropic said it had not determined that Opus 4 definitely crossed the threshold, but couldn’t rule it out. The measures focus narrowly on chemical, biological, radiological and nuclear weapons, plus extra protection against theft of the model’s weights.

Then the ladder itself changed. Version 3.0 of the policy (February 2026) was a comprehensive rewrite. The Centre for the Governance of AI describes it as dropping the old escalating tiers: it specifies mitigations for the next capability level only, with no ASL-4 standards. Anthropic says its ASL-3 protections stay in place. GovAI and other reviewers noted v3.0 also dropped the earlier commitment to pause if safeguards couldn’t keep up. Anthropic’s September 2026 system card for Claude Mythos 5.1 reports against named thresholds instead of ASL numbers: it judges the model has “CB-1” capabilities (it could meaningfully help someone with a basic technical background make a known chemical or biological weapon) but falls short of “CB-2” (replacing the rare expert talent needed for novel weapons).

OpenAI’s Preparedness Framework and Google DeepMind’s Frontier Safety Framework

OpenAI’s Preparedness Framework (version 2, April 2025) tracks three categories: biological and chemical, cybersecurity, and AI self-improvement, each with a High and a Critical threshold. In July 2025, OpenAI said it was treating ChatGPT agent as High in biology as a precaution. In September 2026, its system card for GPT-6 Astra stated that Astra is “our first model to reach the Critical level of cybersecurity capability”, treated it as High (not Critical) for biology, and found it below High for AI self-improvement. It was released with added safeguards and access controls rather than held back.

Google DeepMind’s Frontier Safety Framework (first published May 2024, now version 3.1 from April 2026) is built on Critical Capability Levels (CCLs), with earlier “alert thresholds” and, since v3.1, lower “Tracked Capability Levels”. In its October 2026 safety report for Gemini 4 Argon, DeepMind wrote that it “cannot rule out” that the model has reached the cyber CCL and the biological CCL, and that it built safety cases and mitigations in response.

Strengths. The only ladders here that trigger concrete actions (classifiers, access controls, security rules), tied to specific threats with published evaluations.

Criticisms. The companies write, apply and grade them. Thresholds get redefined between versions, so “still below level X” is hard to check over time. And “High”, “CCL” and “CB-1” can’t be compared directly.

Where AI sits in October 2026. At or near the first “serious” rung everywhere: ASL-3 protections active since May 2025, CB-1 for Anthropic’s newest model, Critical in cyber for GPT-6 Astra by OpenAI’s own rating, and DeepMind unable to rule out two CCLs for Gemini 4 Argon. The labs’ own risk ladders have moved faster in the last 18 months than any of the capability ladders above.

Putting them side by side

Read down the last column: the capability ladders are stuck on arguments about definitions; the risk ladders are moving, because each lab must make a deployment decision and write down why.

And capability and autonomy are separate dials, so the same model can sit in very different places depending on setup.

Where my own research sits

My small research project, RIAI, is a useful test case precisely because it’s small (background: Toward a Terminal Agent That Writes Its Own Skills). Its main line is a terminal agent that learns what commands do by experimenting, rather than being told. In its latest reported run it understood 35 of 39 commands, predicted their effects with 98% accuracy, and passed a 20-task exam with zero calls to a language model. When no command can do a job, it gets a new Python tool written, tests it in a sandbox, and adds it as a new command. But Claude Haiku writes those tools, not the agent itself. And results are checked by the host through a separate channel, never by the agent grading itself.

On the autonomy ladder, inside its sandbox it sits high: it plans and acts on its own, and a separate checker confirms results afterwards, closer to Feng’s Approver or Observer than Operator. But that’s autonomy in a box, on small tasks, with nothing outside at stake. On DeepMind’s grid, it isn’t on the general column at all: it’s narrow, working in one environment, and the part that would make it more general, writing its own code, still depends on a language model. That’s a placement for a research project, not a claim about frontier AI.

How to read a “level” claim

Next time you see “Level 4 AI” or “AGI-level”, run through this:

1. Which ladder? Capability, generality, autonomy or risk? If the claim doesn’t say, the number is meaningless.
2. Who defined the levels, and is it published? A peer-reviewed paper, a published company policy, or a press report of an internal slide? OpenAI’s five levels are the last kind.
3. Measured how, on what tasks? “Expert level” on a coding benchmark isn’t expert level in general. Ask for the test and the date.
4. Capable of, or allowed to? An agent that could run alone isn’t the same as one that does. Autonomy is a setting.
5. Who says so? The company selling it, an independent evaluator like METR or ARC Prize, or a safety framework the company grades itself on?
6. Is the ladder still the same? Frameworks get revised. Anthropic’s ASL ladder was rewritten in 2026, DeepMind’s renamed a level in 2025, and OpenAI’s was never fixed in public.

Key ideas

• There is no single ladder of AI. At least six are in common use, measuring breadth, kind of mind, kind of work, performance, autonomy and risk.
• Narrow / general / super is the oldest and coarsest. Today’s frontier models are clearly beyond narrow; whether they’re “general” depends on the definition.
• The four types of AI (Hintze, 2016) is popular but loose. Language models already blur the line at “theory of mind”.
• OpenAI’s five levels were reported by Bloomberg, never formally published. By most readings, AI is at Level 3, agents, with Level 4 disputed.
• DeepMind’s Levels of AGI separate performance, generality and autonomy. Its 2023 placement of chatbots was Level 1; no official re-rating has been published.
• Autonomy is a dial someone sets. The Feng, McDonald and Zhang levels define it by the human’s role, the same way SAE levels describe driving features.
• The risk ladders are the ones moving fastest. ASL-3 protections since May 2025, OpenAI’s first “Critical” cyber rating in September 2026, and DeepMind unable to rule out two CCLs in October 2026.

Sources: John Searle, “Minds, Brains, and Programs” (Behavioral and Brain Sciences, 1980); Superintelligence (Wikipedia), on Bostrom’s 2014 definition; Tim Urban, “The AI Revolution: The Road to Superintelligence” (Wait But Why, 22 January 2015); Safe Superintelligence Inc. (Wikipedia); NBC New York (CNBC), Zuckerberg memo announcing Meta Superintelligence Labs (June 2025); Hendrycks et al., “A Definition of AGI” (2025); Arend Hintze, “Understanding the four types of AI, from reactive robots to self-aware beings” (The Conversation, November 2016, republished by Robohub); Strachan et al., “Testing theory of mind in large language models and humans” (Nature Human Behaviour, 2024); Axios, “OpenAI nears ‘reasoning’-capable AI” (15 July 2024), reporting Bloomberg; Maginative on Altman at T-Mobile Capital Markets Day (September 2024); OpenAI, ChatGPT agent system card (July 2025); Morris et al., “Levels of AGI for Operationalizing Progress on the Path to AGI” (ICML 2024, v5 September 2025); SAE, “SAE revises levels of driving automation” (2021); Kevin Feng, David McDonald and Amy Zhang, “Levels of Autonomy for AI Agents” (Knight First Amendment Institute, July 2025); METR time horizons and METR on Claude Mythos Preview (2026); Anthropic, “Anthropic’s Responsible Scaling Policy” (September 2023); Anthropic, “Activating AI Safety Level 3 protections” (May 2025); Anthropic, Responsible Scaling Policy v3 (February 2026) and RSP updates; Centre for the Governance of AI, “Anthropic’s RSP v3.0: How it Works, What’s Changed”; Anthropic, System Card: Claude Fable 5.1 and Claude Mythos 5.1 (1 September 2026); OpenAI, Preparedness Framework v2 (April 2025); OpenAI, GPT-6 Astra system card (3 September 2026); Google DeepMind, “Strengthening our Frontier Safety Framework” (September 2025); Google DeepMind, Gemini 4 Argon Safety and Security Report (October 2026).

Related: What Does “AGI” Even Mean?, What an “AI Agent” Actually Is, Without the Hype, and AGI: Where We Are Now (October 2026).