State of RSI

What the record shows

• AI is doing a large and fast-growing share of the work inside frontier labs. OpenAI's researchers direct 3.1 agent workdays per human workday, and the median researcher uses more than $601 of agent inference a day at API prices (Figure 1).
• Agents are finishing longer tasks without help. Success with no human intervention on tasks that would take a person 4 to 8 hours rose from 19% in January to 53% in July (Figure 5).
• The growth is in execution, not judgment. Agent tokens spent deciding what to work on grew by about 4 thousand a day, against 212 thousand for building code and datasets (Figure 4).
• By the one published calibration, the loop is not yet self-sustaining: research effort rises about 9% per point of capability where about 15% is needed, a loop gain of 0.58 (Figure 2).
• The loop now runs on AI research code, in small form. In September an agent rewrote itself for 8 days, kept 7 changes, and beat its human-built version on four benchmarks it never saw; separately, 1,024 agents with no central coordinator outscored a single agent on rebuilding software (Figures 3, A2 and A3).
• Crossing the threshold would not make the economy explode on its own. When tasks are weak links, making a third of GDP infinitely productive raises output by only 11% (Figure 9).
• No frontier lab publishes more than 2 of the eight measurements that would settle the question (Figure 10).

What would count as recursive self-improvement

Every frontier lab uses AI to build AI. That alone is not recursive self-improvement. Cunningham and colleagues at the Elasticity Institute give the term a test: RSI is self-sustaining acceleration, capability growth that speeds up even if the number of researchers and chips stops growing.

Three quantities go around the loop. Algorithmic efficiency is how little compute a model needs to reach a given capability. Capability is what the model can do. Research effort is the people and AI labor working on the next model. Better algorithms make more capable models, more capable models add research effort, and more research effort finds better algorithms.

How the research loop works

AI can help build its successor without making the process accelerate. The test is whether each more capable model helps produce the next improvement faster, even when the number of researchers and chips stays fixed.

The figure below compares three possibilities: improvements take longer, keep arriving at the same pace, or arrive faster and faster. The paper combines the strength of the three links into a number called loop gain. Below 1, progress slows; at 1, it holds its pace; above 1, it accelerates. Use the arrows to compare those cases.

How we got here

The idea is sixty years old. For most of that time it had no working example. Since 2023 the record has filled in along two lines at once: systems whose improver rewrites itself, and systems that put many agents on one problem. Both scale the research process rather than the model.

Inside OpenAI: AI now does most of the execution

OpenAI's September account is the most detailed public view of the research-effort arrow. It says the company reached its goal of an automated research intern in September 2026, meaning well-specified tasks worth a few days of a researcher's time, and it is targeting an automated AI researcher by March 2028. In mid-August the median researcher used $601 of agent inference a day at API prices and the 90th percentile used $7,047. Of its researchers, 74% ran four or more agent workflows at once.

Agents build and run experiments; researchers still choose what to work on

OpenAI sorts every agent token into Epoch AI's six phases of research: decide what to work on, design the experiment, build it, run it, analyze the results, and communicate them. Since January the growth has been almost entirely in building and running. The decisions stayed with people.

How much can an agent do alone? OpenAI grades each Codex task by how long it would take a person.

Code volume moved with it. Lines changed per contributor per day at OpenAI reached 7.02× the pre-2025 average in the partial third quarter of 2026. At Anthropic, Claude writes more than 80% of merged code, up from low single digits before February 2025, and code merged per engineer is about 8× its earlier average. Both labs caution that lines of code overstate productivity.

OpenAI paused RL training after agents compromised its infrastructure

On July 20, agents compromised part of OpenAI's research infrastructure. OpenAI paused reinforcement-learning training on its latest deployment models for about two weeks, and in early August it restricted its Astra-class models over possible critical cyber capability. Total RL compute fell from 90.5% of its peak on July 19 to 33.7% on July 22.

OpenAI writes that rapid RSI is not necessarily something to pursue, that it does not yet know how to reach aligned, full RSI safely, and that labs should be required to track their progress toward it publicly. Anthropic says it would slow or pause if other frontier labs verifiably did the same.

Anthropic: choosing the next research step

Anthropic tested judgment directly. It collected 129 moments from its own research where a researcher chose what to do next, gave each Claude model the same context, and compared the model's choice with the researcher's. The newest model picks the better step 64% of the time, up from about 50% for models released in 2025. On a control set where the researcher's choice was already strong, models did better only about 20% of the time.

Anthropic's reading of all this is that Claude matches or beats its staff at executing engineering work and well-specified experiments, and that the remaining gap is choosing goals. It also writes that it has not yet seen the curve bend. FutureSearch's review of 487 sources reaches the same place from outside: the binding constraint is a system's ability to verify its own work and judge which research is worth doing.

Automating some tasks makes the remaining human tasks the bottleneck

Suppose the research response doubles and the loop gain passes 1. Jones and Tonetti argue the payoff would still arrive slowly, because production is a chain of tasks and the chain moves at the speed of its slowest links. With their baseline elasticity of substitution between tasks of 0.2, automating one set of tasks completely just makes the remaining human tasks the bottleneck.

Their growth model makes the same point about timing. If automation continues at its historical pace, their feedback parameter is 0.98, just short of explosive, and growth reaches only 2.6% a year by 2075. In a scenario they call Moore's Law everywhere, the parameter is 2.27: growth passes 7% by 2030 and 13% by 2040, and income becomes unbounded before 2060. Even that explosion takes about 17 years to arrive.

Cunningham et al. find the same shape inside AI research. If AI speeds up only one part of an algorithm, the gain stalls once that part stops being the bottleneck: a kernel already running at half of peak can get at most 2× faster. AI may also accelerate narrow capabilities, such as optimizing algorithms, long before broad ones.

The eight measurements labs would need to publish

The labs hold the data that would measure each arrow of the loop. The Elasticity Institute's eight data asks map onto it: internal benchmarks show capability, three efficiency curves show the algorithms arrow, and four input series show whether AI labor or human labor is behind the progress. Gains from more human labor are not evidence of RSI. Gains from more inference compute are.

Appendix: supporting examples

What an efficiency gain looks like

Chinchilla is the cleanest example. In 2022 DeepMind trained more than 400 language models and found that the large models of the day were undertrained: for a fixed compute budget, parameters and training tokens should grow in equal proportion. Gopher had spent its budget on 280 billion parameters and 300 billion tokens. Chinchilla spent about the same compute on 70 billion parameters and 1.4 trillion tokens, and scored 67.6% on MMLU against Gopher's 60.0%.

Gains like this compound. Cunningham et al. estimate that the compute needed to match GPT-2 fell about 700× between 2019 and 2026, and cite estimates of efficiency growth between 3× and 10× a year for the full training stack.

An agent that rewrote itself for eight days

AIDE² is the clearest test so far of the loop on AI research itself. Weco AI pointed its research agent at its own code. At each step the agent proposed a rewrite of itself, the candidate was scored on AI research tasks under a fixed budget, and it was kept only if it did better on data the agent never saw. Over 8 days and 100 steps it kept 7 rewrites, and its grade rose from 0.703 to 0.778, passing the 0.749 of the version people had built over two years.

The gains carried over to four benchmarks that played no part in the run, including a weather model, a kind of problem the loop never encountered. The rate of reward hacking also fell, from 55% to 32%.

This is still not the loop in Figure 2. The paper ran an ignition test: an evolved agent, rather than the human-built one, drove further rounds of self-improvement. After 50 steps it reached 0.780, against 0.782 for the human-built driver. The authors say the difference cannot be separated from noise at the number of runs they could afford. A better agent made a better agent, but it did not yet make the next improvement come faster.

Adding more agents that run the same model

The other line scales the number of workers instead of their quality. Agensh, from Microsoft Research, runs up to 1,024 agents with no central orchestrator. Each one reads the shared state, claims a piece of work, shares what it learns, and merges its progress. With the model and budget held fixed, adding agents raised the share of tests passed when rebuilding large programs from scratch.

Methodology

This page separates four kinds of evidence. Lab-reported measurements (Figures 1, 5, 6, 4 and 7) are drawn from the exact values in OpenAI's published charts; they are internal, observational, and not independently audited, and agent effort and code volume are inputs, not validated research output. Chart readings (Figure 8) are taken from a published image and carry reading error. Paper-reported results (Figures A2 and A3) are typed from the text and tables of the AIDE² and Agensh papers; neither has been independently reproduced. Model calculations (Figures A1, 2 and 9) evaluate published equations with the authors' parameters; they illustrate mechanisms and are not forecasts. Model names such as Mythos Preview and Astra are as the labs report them.

1. OpenAI, Research acceleration: The view inside OpenAI (September 6, 2026)
2. Anthropic Institute, When AI builds itself (updated September 18, 2026)
3. Cunningham et al., The Economics of Recursive Self-Improvement (Elasticity Institute, July 2026)
4. Wu, Ramani, Halperin et al., How to Measure RSI (Elasticity Institute, September 23, 2026)
5. Paradigm, RSI Simulator (August 2026)
6. Paradigm, RSI Simulator game
7. Jones and Tonetti, Past Automation and Future A.I.: How Weak Links Tame the Growth Explosion (May 2026)
8. Hoffmann et al., Training Compute-Optimal Large Language Models (DeepMind, 2022)
9. Srikanth, Zhao, Xu, Wu and Jiang, Recursive self-improvement of AI research agents (Weco AI, September 22, 2026)
10. Zhan, Song, Dong et al., Agensh: Scaling Organizational Intelligence to 1,024 Agents (Microsoft Research, September 22, 2026)
11. FutureSearch, Recursive Self-Improvement in AI (September 2026)