Guide
The AI ladder: how close are we to really smart AI?
Is AI about to match people, or is it overhyped? One simple way to see where things stand is a ladder of five levels, first described by OpenAI. Today's AI has clearly climbed the first two rungs and is partway up the third. Here is what each level means, the milestones that got us there, and what is still missing.
Level 5 Organisations
Not yetAI that can do the work of a whole company: many agents planning, building, selling and managing together.
Still missing: Almost everything. Small teams now build a lot with AI (see our Built with AI cases), but people still set the goals, make the decisions and carry the responsibility.
Level 4 Innovators
Early signsAI that comes up with genuinely new ideas and discoveries, not just remixes of what it has read.
- 2025Google's AI co-scientist proposes drug candidates for a type of leukaemia that lab tests then support. research.google
- 2025DeepMind's AlphaEvolve finds a faster way to multiply some matrices, beating a record from 1969, and saves 0.7% of Google's computing power. deepmind.google
Still missing: Independence. Today's discoveries are narrow and come from systems built for one kind of problem, with scientists choosing the questions and checking every result.
Level 3 Agents
You are hereAI that does multi-step tasks for you: it plans, uses apps and websites, and checks its own work.
- 2024Anthropic's Claude learns to use a computer by looking at the screen, moving the mouse and typing. anthropic.com
- 2025Researchers at METR find the length of tasks AI agents can finish has doubled about every 7 months for six years. metr.org
- 2026Stanford's AI Index reports agents completing 77% of a set of real-world tasks, up from 20% a year earlier. hai.stanford.edu
Still missing: Trust. Agents work well for coding and research, but still make costly mistakes, can be tricked by hidden instructions, and need a person checking the important steps.
Level 2 Reasoners
ReachedAI that works through hard problems step by step, like a skilled person with a pen and paper.
- 2024OpenAI's o1, the first widely used reasoning model, "thinks" before it answers. openai.com
- 2025An AI model from Google DeepMind reaches gold-medal standard at the International Mathematical Olympiad. deepmind.google
Still to fix: Reliability. The same model that solves an Olympiad problem can still slip on a simple puzzle or a long, messy real-world task.
Level 1 Chatbots
ReachedAI you can talk to in everyday language, which answers questions, writes and explains.
- 2022ChatGPT launches, and millions of people start talking to AI for the first time. openai.com
- 2026ChatGPT alone reaches 900 million weekly users. techcrunch.com
Still to fix: Nothing big: chatbots are everyday tools. They still make things up at times, so important answers need checking.
Meanwhile, AI learned new senses
Alongside the ladder, the same models went from text only to seeing, hearing and making pictures and video.
- Text2022Chat, write and explain
- Images2023Understand and create pictures
- Voice2024Talk back and forth in real time
- Video2025Make realistic video clips
Where the five levels come from
In July 2024, OpenAI told its staff it would track progress towards AGI on a five-level scale, from chatbots to AI that can run an organisation. Bloomberg reported it first. It is one company’s view, not an official standard, and experts argue about where the lines fall. Google DeepMind proposed a different scale in its “Levels of AGI” paper, which grades AI by how well it performs compared with skilled people and by how widely it can do so. When it was published in 2023, the paper rated chatbots such as ChatGPT as “emerging” general AI, the first of its five levels. Both scales tell a similar story: real progress, with the hardest steps still ahead.
How we judge each level
“Reached” means the ability is widely available and works most of the time. “You are here” means it works in real use, but not reliably enough to leave alone. “Early signs” means a few impressive results in narrow areas. These are our judgements, based on the sourced milestones on each rung, and we review them every quarter because things move fast. For the full story so far, see the AI timeline, and for the numbers, where AI stands in 2026.