What is superintelligence, and how close is AI to it in 2026?

In September 2026, superintelligence stopped being a word from philosophy books and became a headline. An AI lab published a claimed solution to a famous unsolved maths problem, US lawmakers introduced a bill to ban superintelligence outright, and a group of researchers from the biggest labs warned about an “intelligence explosion”. If you build software for a living, as I do, it is worth separating what is actually happening from the noise.

What is superintelligence?

The term was popularised by philosopher Nick Bostrom in his 2014 book Superintelligence. In his framing, a superintelligence is any mind that far outperforms humans across almost every area that matters: science, strategy, engineering, persuasion, planning. Not better at one game or one benchmark. Better at nearly everything.

That is the key difference from the AI most people use today. A model that writes excellent code but cannot decide which product to build is powerful, but it is not superintelligent.

AGI vs superintelligence

The two terms get mixed up constantly, so here is the simple version:

• Narrow AI is very good at a specific task: translation, image recognition, recommending products.
• AGI (artificial general intelligence) roughly matches a capable human across most cognitive work. Definitions vary a lot, which is part of why forecasts disagree.
• Artificial superintelligence (ASI) clearly exceeds the best humans across most domains, including the hardest one: deciding what is worth working on.

Many researchers think the jump from AGI to superintelligence could be short, because an AI that is as good as a human AI researcher could be copied thousands of times and pointed at improving itself. That idea is called recursive self-improvement, and I cover it in detail in how AI is starting to build AI.

Where AI actually is in 2026

Three data points show how fast things are moving.

Tasks are getting longer

The research group METR tracks how long a task an AI agent can complete on its own. According to Anthropic’s analysis of that data, this length has recently been doubling about every four months, faster than the earlier seven-month trend. Anthropic reports its models went from roughly four-minute software tasks in early 2024 to tasks of around twelve hours two years later.

AI is writing most of the code at AI labs

Anthropic says that by May 2026 more than 80% of the code merged into its codebase was written by Claude, up from low single digits before its coding agent launched in early 2025. A later report put AI’s share of the company’s research and development work at 26% by August 2026. The humans still choose the direction; the models increasingly do the execution.

Frontier maths results

On 8 September 2026, OpenAI published a claimed resolution of the Navier–Stokes problem, one of the Clay Institute’s Millennium Prize Problems, produced by thousands of AI agents running on an unreleased model and accompanied by a formal proof in Lean. It is an extraordinary claim, but it has not yet been independently verified, the Clay Institute still lists the problem as open, and it came with a public dispute over credit involving outside researchers. Treat it as a strong signal, not settled history.

When will superintelligence arrive? What forecasters say

Nobody knows, and anyone giving you a precise date is selling something. What we do have:

• Forecasting communities such as Metaculus put their median for AGI in the early 2030s, and aggregated forecasts across platforms cluster around 2031, with a wide range stretching into the 2040s.
• Lab leaders tend to give shorter timelines. Keep in mind they also raise money and recruit on those narratives.
• Forecasters have repeatedly pulled their dates earlier over the past three years, and in early 2026 several well-known forecasters who had pushed their dates out in 2025 moved them closer again.

The first question to ask about any timeline is which definition it uses. “AI that can do most knowledge work” by 2028 and “superintelligence” by 2028 are very different claims.

The case for caution: it might be slower than the hype

There is real evidence on the other side. A Princeton-led study covered by MIT Technology Review found that AI agents can handle the engineering side of AI research but still lack the judgement and originality to produce top-conference research on their own. Anthropic’s own essay makes the same point: models now match or beat skilled humans at running a well-defined experiment, but they still lag at choosing which problems matter.

Physical limits matter too. More capable AI needs more chips, more data centres and more electricity, and those grow at the speed of construction, not software. Many of today’s exponential curves may turn out to be S-curves.

Superintelligence risks and the policy fight

The risks people worry about fall into three groups:

• Loss of control. If an AI system can design its successor, small alignment problems could compound in ways humans cannot inspect.
• Misuse. Extremely capable systems could supercharge cyberattacks, surveillance or large-scale manipulation.
• Concentration of power. If only a few companies or governments control the most capable systems, they gain enormous economic and political leverage.

Politics caught up this month. US Senator Bernie Sanders and Representative Greg Casar introduced a bill that would ban the development of artificial superintelligence and push for a global ban; groups like MIRI endorsed it, while critics argue a unilateral ban would simply hand the lead to other countries. At the UN General Assembly, President Trump announced the US would start calling AI “super intelligence”. And on 28 September, research leaders from OpenAI, Anthropic, Microsoft and Meta published a paper urging oversight of self-improving AI before any acceleration begins.

Supporters of a slowdown say the risks are too large to gamble on. Opponents say slowing down in one country only helps less careful competitors, and that the benefits for science and medicine are too big to delay. Both sides agree on one thing: the decisions are being made now, not in some distant future.

What superintelligence means for your business today

You do not need to pick a side in the timeline debate to act sensibly. Whatever happens with superintelligence, the capabilities that already exist are enough to change how small teams work:

• Automate execution, keep judgement. Today’s models are excellent at doing well-specified work and weaker at deciding what matters. Design your workflows the same way. I explain how in agentic AI for business.
• Build for model swaps. Capabilities jump every few months. Keep your prompts, tools and data separate from any one model so you can upgrade without a rewrite.
• Know your costs. Agents that run for hours consume a lot of tokens. Plan the budget before you scale; see how much an AI agent costs to build and run.
• Keep humans in the loop where mistakes are expensive. Payments, legal text, anything customer-facing.

Frequently asked questions

Does superintelligence exist yet?

No. Today’s systems are superhuman at some narrow tasks and increasingly strong at long, well-defined work, but they are not better than the best humans across nearly all domains.

Is AGI the same as superintelligence?

No. AGI roughly matches human ability across most cognitive work. Superintelligence clearly exceeds it. Many researchers expect the gap between them to be short once AGI exists.

Which companies are trying to build superintelligence?

OpenAI, Anthropic, Google DeepMind and Meta all work on frontier models, and some organisations say it openly: Meta has a division called Meta Superintelligence Labs, and Ilya Sutskever’s company is called Safe Superintelligence Inc.

Should I be worried?

Informed, yes. Paralysed, no. The practical move for most businesses is to use today’s AI well, keep control of your own data and systems, and follow the policy debate, because it will shape the tools you can use.

If you want to put today’s AI to work safely inside your own systems, see my agentic systems and MCP server development service or get in touch.