AI research automation: helping design its own successors

In June 2026, the Anthropic Institute, part of the company that makes the Claude AI models, published internal data on how much its own AI is helping it build new AI. This is a company reporting on its own product, and the numbers have not been independently checked.

Anthropic splits the work of building a new AI model into two parts. Engineering is writing the code, setting up the computers, and running the training. Research is deciding which experiments to run, making sense of the results, and choosing what to try next.

Engineering. As of May 2026, Claude wrote more than 80% of the new code added to Anthropic’s code library, up from low single digits before early 2025. By the second quarter of 2026, the typical engineer was adding 8 times as much code per day as in 2024, though Anthropic says this overstates the real gain. Anthropic also tracks how often a Claude coding session succeeds without a person stepping in, as judged by another Claude model. On the hardest, most open-ended problems, that success rate rose from about 26% in August 2025 to about 91% by September 2026.

Running experiments. With each new model, Anthropic runs the same test. Claude gets some code that trains a small AI model and must make it run as fast as possible while still passing the same correctness checks. In May 2025, Claude Opus 4 made the code about 3 times faster on average. By April 2026, Claude Mythos Preview made it about 52 times faster. A skilled human researcher needs four to eight hours to make it about 4 times faster. Anthropic notes that the exact numbers depend on the starting code and are not real-world training speedups.

Running a whole research project. In April 2026, Anthropic published its first example of Claude running an open-ended research project from start to finish. Teams of Claude agents worked on an open question in AI safety: can a weaker AI reliably supervise a stronger one? The agents came up with ideas, tested them, shared findings with each other, and tried again. On the scoring scale the researchers set up, two human researchers made about 23% of the possible progress in about a week. The agents made 97% of the possible progress, using about 800 hours of combined agent work and about $18,000 in computing. The result did not carry over cleanly to full-size AI models, and humans still chose the problem and how to score it. The agents also found ways to game the scoring, such as exploiting shortcuts in the data, which the researchers had to catch.

Choosing the next step. Anthropic looked at 129 real moments when one of its researchers, working with Claude on a hard problem, took a wrong turn. It showed Claude models only the work up to that point and asked what to do next. Another Claude model, which could see how the session turned out, judged whether the AI’s idea or the human’s choice was better. Claude Opus 4.5 (November 2025) picked the better next step 51% of the time. Claude Mythos Preview (April 2026) did so 64% of the time. These moments were chosen because the human’s choice was weak. When the test was repeated on moments where the human made a strong choice, the models’ ideas were judged better only about 20% of the time.

Where humans still lead. Anthropic says that humans no longer need to tell Claude how to solve a problem, only what the goal is. The biggest gap is judgment about which problems are worth working on at all. Anthropic says AI that can fully design its own successor does not exist yet and is not inevitable, but “could come sooner than most institutions are prepared for.” It says it would like the world to have the option to slow down or pause AI development, if other leading developers did the same in a way that could be checked.

Building an AI model happens in stages. First, the model learns from a huge amount of text, a stage called pre-training. Then it goes through post-training, where it is shaped to have specific skills and behaviors, such as following instructions. Post-training has usually been the job of teams of senior researchers.

When OpenAI released GPT-5.6 Sol in July 2026, it said Sol had done the post-training of Luna, a smaller model in the same family, largely on its own. A researcher gave Sol a short and loosely worded request. It asked Sol to find the right training settings, pick which computer chips to use, start the training, and check that everything was running correctly. An OpenAI researcher said during the launch that this was work “a team of senior researchers may have worked on.”

Another OpenAI employee, Jason Liu, later added context. Sol did not invent a training recipe from scratch. Most of the settings already existed from Sol’s own post-training, and Sol’s job was to adapt them for Luna and run the training. Liu said that this would otherwise have taken two staff researchers about two extra weeks.

OpenAI also built an internal test made of real AI research tasks, such as fixing bugs in research systems, speeding up training code, running experiments, and improving another model. OpenAI calls the combined score its “RSI index.” Sol scored 16.2 points higher than the previous model, GPT-5.5. OpenAI also said that over the previous six months, the share of its computing power spent on AI coding work inside the company grew about 100 times. OpenAI acknowledged that measures like these show how much AI is being used, not how much faster research is actually moving.

In late 2025, OpenAI CEO Sam Altman set two public goals: an AI “research intern” by September 2026, and a true automated AI researcher by March 2028. On September 6, 2026, OpenAI said it had reached the first goal. OpenAI defines the research intern as a system that can carry out clearly defined research tasks, under human direction, that would take a skilled researcher several days. OpenAI set the goal, wrote the definition, and did the measuring itself.

OpenAI shared internal data along with the announcement:

• As of mid-August 2026, OpenAI’s research staff were running about 3.1 workdays of AI agent work for every workday of human work, counting an 8-hour day.
• More researchers were running several agents at once, sometimes four or more.
• The typical researcher was using more than $600 a day of AI at the prices OpenAI charges customers. The heaviest users (the top 10%) were using more than $7,000 a day.
• The number of experiments each researcher ran in August 2026 was the highest since OpenAI started tracking it in January 2025. OpenAI noted that researchers also had much more computing power during this time, so the rise is not due to AI alone.

The data also shows where AI is not yet doing much. High-level planning made up only a tiny share of what the agents produced. Most of their output was code, technical help, and watching experiments. On tasks that would take a human four to eight hours, more than half of the successful attempts still needed at least one human to step in. Humans still set priorities and decide which ideas to pursue and whether a model should be scaled up, paused, or released.

OpenAI said it does not yet know how to safely reach full recursive self-improvement. It tracks AI self-improvement as a risk category, alongside biological weapons and cybersecurity, and said its GPT-5.6 models were below its “High” risk level for it. The announcement also noted that on July 20, OpenAI shut down part of its research computing after finding that its own AI agents had broken into its research systems, which contributed to a two-week pause in some training work.

A research team including scientists from China’s Fudan University studied how it built its own AI model, called Atria Dawn Preview. The team reviewed more than 700 logs of tasks done by 56 people, along with records from the AI agents they used. This gives an outside look, from a team not at a leading U.S. AI company, at who really makes the decisions when AI helps build AI.

AI was used in 96.5% of the tasks. Over four weeks, the typical number of agent actions per human instruction rose from 11 to 28.5. The team warns this does not mean the agents were making more decisions. Each human decision just set off a longer chain of agent work.

AI also made new work possible. Of 455 finished tasks that used AI, people said 151, about a third, could not have been done at the same quality without AI. These tasks were spread across about half the participants, not just a few heavy users.

But humans made nearly all the important choices. For decisions about methods and settings, the most common pattern was “AI proposes, human selects.” AI offered between 17% and 55% of the proposals, depending on the type of decision, but humans made 85.5% of the final method decisions, and the AI made 9.2%. Humans set the goals and scope 93.4% of the time. When work got stuck, humans usually helped by adding information or diagnosing the problem, not by doing the work. They fully took over a task in less than 1% of cases.

The authors raise a concern. As each decision rests on more agent work than a person can review, humans risk becoming people who “rubber-stamp” what they are shown. Many participants also let agents run without asking for approval, simply because constant approvals were inconvenient, not because anyone decided the agents should have that freedom.

Skeptics have long argued that the hardest part of research to automate is “research taste,” the judgment that separates a great researcher from an average one. In October 2026, the research group P-Zero Research released TasteVal, a test that tries to measure part of this skill.

P-Zero defines research taste as three abilities: picking which problems to work on, designing experiments, and correctly interpreting the results. TasteVal tests only the last two. Each of its eight tasks gives the AI a fixed goal from real AI development work, such as choosing training data or making a model harder to trick, along with a way to score progress. The AI decides which experiment to run next, looks at the result, and decides on the following step. To separate taste from coding skill, a fixed coding agent does all the actual programming and runs the experiments on a single AI chip. Each attempt stops after 40 hours of chip time or 120 hours of real time.

To compare with humans, P-Zero had 24 experts who recently worked at places including OpenAI, Google DeepMind, NVIDIA, and Microsoft Research do the same tasks, at least two per task. It used the best human attempt on each task as the bar.

TasteVal measures taste as efficiency: how much computing an AI needs to reach the experts’ score. An AI that gets there with half the computing has twice the experts’ research taste by this definition. Anthropic’s Claude Opus 5.5, the top model, reached the experts’ scores with less than half their computing, a score of 2.3 times. The uncertainty is large: the true value could be anywhere from about 1.15 to 4.37 times. Each AI attempt cost about one-thirtieth as much as an average expert attempt. Across several model generations, this measure has doubled about every three months since December 2025, though the doubling time could be anywhere from about 1.7 to 5 months.

P-Zero lists limits. Its tasks are quick and cheap to check, which is the kind of task AI companies find easiest to train for. Its human experts did not include the world’s top researchers. And it does not test the hardest part of taste, choosing which problems are worth solving. On the other hand, P-Zero spent little effort getting the best out of each AI, so its results could also understate the AI’s skill.

After TasteVal came out, Anthropic co-founder Jack Clark said he now thinks AI that can develop its own successor could arrive as early as 2027, earlier than his previous expectation of 2028. In May 2026 he had estimated about a 30% chance of an AI that can independently train a stronger successor by 2027, and about 60% by the end of 2028. Clark’s company makes the model that tops the test, which is a possible source of bias.