A new Anthropic document shows a sharp jump in Claude's abilities in writing code, fixing systems, and research development — and raises the question of whether humanity can slow down the pace of progress of advanced AI systems over time.
Anthropic published last week A document with some predictions Which should worry – and delight – us all. Their less extreme prediction is that by the end of the year, artificial intelligence will be able to write better code than a human software developer.
This prediction in itself is a dramatic change for programmers, who will see a large portion of their workload shifted to AI. It also has great significance for all the other ordinary people who don't know how to code, as each of us will have a "personal genius" who can develop any application we want.
But this is the least extreme forecast.
The real news that Anthropic announced is that if technological trends continue to advance at the same pace, it is likely that artificial intelligence systems will be able to plan, design, and improve themselves – and they will do so faster and better than any human working on those systems.
To understand why Anthropic thinks this is what will happen, we first need to look at the evidence it brings from the field – that is, from the ways in which Anthropic engineers themselves use artificial intelligence.
As of May 2026, Claude wrote more than eighty percent of the code that found its way into Anthropic's products. To put that into perspective: at the beginning of 2025, the rate of code written by Claude was only around five percent.
Of course, Claude receives constant guidance from the engineers and software developers who run it. But if they do it right, it empowers their capabilities. The average engineer at Anthropic now produces eight times more code every day than he did just two years ago. That engineer’s job has become a management job: he directs the AI and tests its products, rather than having to write the code himself.
"We don't reward people based on the number of lines of code they write," Anthropic explains, "but team members produce more simple code because they use artificial intelligence systems."
According to an internal survey of the company's research teams, engineers claim they are achieving four times greater productivity than they could without the use of artificial intelligence. They may well be right.
“In April 2026, Cloud launched more than 800 code fixes, reducing API errors by a factor of 1,000,” the document states. “The engineer running Cloud estimated that it would have taken four years of human labor to complete the task.”
A year ago, Claude could already write high-level code… but only for ‘trivial’ and simple problems. More complex problems, especially open-ended problems that require creative thinking, planning ability, and a broader understanding of the domain, were left far behind. By the end of 2025, Claude had reached only a twenty percent success rate in solving such complex problems.
But that was six months ago.
The past six months have seen a meteoric rise in Claude’s ability to solve open-ended problems. In fact, the improvement in capabilities is the kind we would expect to see over a decade or two. Today, in early June 2026, Claude has already reached a 76 percent accuracy level in solving complex problems that only a human engineer could consistently tackle in the past.
What does this look like in practice?
"A routine system upgrade began to cause tens of thousands of training jobs to crash." Anthropic cites an example from the field. "An engineer directed Claude to a real-time event... Claude went through the running jobs and tested one runtime environment at a time, isolating the single, hidden signal that caused the crash, reproducing it consistently, and approving a fix for it. In about two hours, Claude delivered a product that would normally take him two to three days of work."
The two pieces of evidence I presented explain why Anthropic believes that if the improvement trend continues at the same pace, then by the end of the year, Claude will be at a higher level than the human programmers in society in dealing with problems of all kinds.
In fact, it could be argued that Claude has already surpassed human engineers in identifying problems, at least. At Anthropic, they developed an automated tool that scans all new code for bugs, security flaws, and other flaws. They ran it retroactively—that is, on code that human developers had written in the past, and only after launch did bugs become apparent. Claude found about a third of the bugs.
"The engineers who wrote that code are some of the best in the world at building these systems," Anthropic wrote. "Claude is now catching the mistakes they missed."
At Anthropic, there's a test they like to do every time they release a new Claude model: they give Claude a small artificial intelligence trainer, and ask him to make the code run as fast as possible, but without compromising the quality of the products.
In May 2025, Cloud Opus 4 managed to make the code run three times faster. This is a level of success similar to that of a skilled human researcher, who accelerated the code four times when asked to do a similar task.
In April 2026, Claude Mythos (Mythos Preview) improved the code execution speed by 52 times. You read that right. And Anthropic understands the meaning, and wrote that –
"In this part of the study – improving steps in a well-defined experiment – Claude went from "very helpful" to superhuman in less than one year."
Every researcher knows how important intuition – the sudden understanding of where to go from here to solve a problem – is in research and development work. At Anthropic, they decided to test whether their models develop similar intuition. They took a set of open problems that their researchers had worked on in the past, and identified the moment when the human researcher made a mistake and moved to explore the wrong direction for the solution, before eventually ‘recovering’ and getting back on track. Anthropic showed Claude the work up to the point where the human researcher got confused, and asked him what he suggested doing next.
Anthropic’s best model six months ago – Opus 4.5 – provided better answers than humans 51 percent of the time. Since then, the models have improved even more, and Claude Mitus is already able to outperform human intuition 64 percent of the time.
To clarify what this means: Claude can already begin to offer directions of thought at a level that is close to that of the human, and perhaps (just perhaps) surpasses it.
This does not mean, of course, that it does not have mistakes. And as is the way with artificial intelligence today, some of its mistakes will be perceived by us as stupid, delusional, and ridiculous. But in a broader perspective, the correct use of artificial intelligence in research and development will lead to more correct decisions, already in the near future – that is, this year.
When you take all this evidence together, you can understand why Anthropic believes that in the coming years, artificial intelligence will be able to reach a level where it will design and run experiments on its own, and will even be able to improve itself from generation to generation – without human intervention.
Anthropic themselves admit that they are not entirely sure that we can easily reach this level. The most obvious point of contention is that deciding which problems to work on still limits Claude. It is not obvious that today's AI systems will be able to cope with the need for a broader systemic and strategic vision. But even if AI can automate 'only' much of the Sisyphean and tedious work of conducting experiments and gradually improving systems, it is still expected to leap forward science, technology - and also the pace of its own development.
And that is the most conservative and cautious scenario.
In the more likely short-term scenario, Anthropic predicts that AI will continue to improve in its capabilities – but humans will still be the ones defining the directions of research and deciding which products are better or worse. If this is what happens, then organizations (and individuals) that use AI effectively will be able to produce results a hundred, or a thousand, times more efficiently than they are today. This is the scenario Anthropic believes we are moving towards.
But there is an even more extreme scenario.
In the most extreme scenario presented by Anthropic, AI begins to improve itself without human involvement or a real ability for humans to understand the process in depth. Engineers will mainly try to monitor these systems, test them, and verify that they do what is expected of them and do not get out of control.
For this to happen, technological trends need to continue to advance at the same pace as they have in recent years, and AI systems need to develop intuition and problem-solving abilities on a level similar to humans. Based on the evidence Anthropic has presented so far, it seems that we are indeed safely on the path to realizing this scenario – although probably not in the next year or two. Probably.
What would a world look like in which artificial intelligence plans research and products, and also conducts the research required to implement them?
First of all, this is a world where computing power – in English, compute – determines everything. Those who have more computing power in their hands can produce more advanced products, faster.
Second, as I have been saying for several years now, it will be a world of miracles and wonders. Artificial intelligence will be able to conduct research in all fields – including biology, chemistry, physics and so on. It will also be able to build on its results from previous experiments and move forward rapidly. This means that we will see medical, engineering and mathematical developments that come almost ‘out of nowhere’. Recent breakthroughs in mathematics – with 17 Erdoğan problems solved since the beginning of 2026 alone – hint at what such a world will look like. We will see solutions to large and complex problems in all fields, coming year after year.
At the same time, the risk that we will lose control of the systems that conduct this research will increase every year. When we do not understand how they work, and find it increasingly difficult to follow the 'thinking' that guides them, they may get out of control and make decisions that are not necessarily in the best interests of humanity.
And here comes Anthropic's final point: that the progress of artificial intelligence needs to be stopped.
Yes, seriously.
Anthropic unequivocally states that they would prefer to stop the development of technology – the same technology that has made them one of the most important and respected companies in the world.
“We think it would be a good thing if the development of the technology could be effectively slowed down,” they write, “to give ourselves more time to deal with its enormous implications.”
The trouble is that it is not clear how to make sure that everyone slows down at the same time. Anthropic is not run by little children, and they understand very well that in the current arms race between the superpowers and societies, if only one side stops, then the other side will continue to run forward with all its might.
Anthropic's proposal is to set up systems to monitor the world's leading laboratories to ensure that they have truly slowed down their development. Such systems would have to monitor laboratories around the world, and those laboratories would have to agree to stop or slow down their activities collectively.
It's a nice proposal, reminiscent of the international treaties signed to slow the pace of development of nuclear weapons. But can it really be implemented? Anthropic isn't sure, but they're willing to try.
“In the coming months, we will organize conferences where policymakers, researchers, gatekeepers, and other AI companies can help answer some of the questions raised in this report,” they write at the end of the document, “particularly regarding full recursive self-improvement, and how to create better options for coordination and collaborative thinking. We will publish what comes out of these conferences.”
It appears that we are entering the final stretch before reaching the “intelligence explosion” – a term coined by Irving John Goode sixty years ago. This will be the time when artificial intelligence will be able to improve itself on its own, and “human intelligence will be left far behind.”
Artificial intelligence researcher Vernor Wing stated that "shortly thereafter, the age of humans will end."
What will come next? It's hard to know. But Anthropic is well aware of the responsibility that falls on their shoulders.
“The window of time to explore these questions together is here,” they write, “and people outside of AI companies need to be involved in the discussions.”
Short FAQ
What does Anthropic claim about Claude?
According to the text, Anthropic shows a sharp improvement in Claude's abilities in writing code, fixing bugs, improving performance, and suggesting research directions.
What is recursive self-improvement?
This is a scenario where an artificial intelligence system not only assists engineers, but participates in the design and improvement of the next generations of AI systems.
Why does Anthropic want to slow down development?
The fear is that AI systems will advance faster than humans can monitor, understand, and control them, especially if several companies or countries continue to compete with each other without coordination.
What does this mean for programmers and researchers?
In the short term, this means moving from manually writing code and experimenting to managing, directing, and controlling AI systems. In the longer term, it could mean a profound shift in the way research and development is conducted.
More of the topic in Hayadan:
