Key takeaways
- Recursive self-improvement is AI helping build better AI in a loop that speeds up; no lab says full RSI has happened.
- Anthropic's November 9, 2026 system card puts its internal acceleration at about 1.88x, below its 2x line, and calls that uncertain.
- Anthropic says it limited some AI-research skills in Fable 5.2, holding kernel and architecture work near Opus 5.5 levels.
Recursive self-improvement, or RSI, is when an AI system gets good enough at AI research that it helps build a better successor, which then helps build a better one after that, so progress speeds up on its own. It is the idea behind a phrase now in the news: on November 9, 2026, Anthropic released Claude Fable 5.2 and Mythos 5.2, and TechCrunch's headline put RSI in the first line.
No lab says full RSI has happened. What the labs now say is that AI already does a large share of the work of building AI, and that they measure how much that speeds them up.
What recursive self-improvement means
The loop has three steps. A model helps write code, run experiments and design training for the next model. The next model is better at those jobs. The gain compounds.
It is not the same as a model that learns during a chat, or an agent that tunes its own prompts. Those are small, bounded versions. The version that worries safety researchers is the one where the loop runs without humans in the middle, because a fast loop could outrun the checks that tell people whether each new model is safe.
Two levels are worth keeping apart:
- Assisted acceleration. Humans still direct the work, but AI does much of the labor. This is where Anthropic says it is today. Its announcement says over 90% of the changes merged into its core codebase are now written by Claude, with human guidance.
- Full RSI. AI does the research with little or no human input. OpenAI's Preparedness Framework calls this "fully automated AI R&D".
How the labs define the threshold
Both labs publish a line they say they will not cross without stronger safeguards. The wording differs.
| Anthropic | OpenAI | |
|---|---|---|
| Policy | Responsible Scaling Policy, version 3.4, effective July 8, 2026 | Preparedness Framework, version 2, April 15, 2025 |
| Lower bar | Models that can fully substitute for its research scientists and engineers at competitive cost (within a factor of five) | "High": the model's impact is equivalent to giving every OpenAI researcher a mid-career research engineer assistant, against a 2024 baseline |
| Higher bar | "Dramatic acceleration" of AI progress for reasons tied to automating AI R&D. The system card reads this as roughly a doubling of internal progress | "Critical": a model capable of "recursively self improving", defined as a superhuman research-scientist agent, or a generational model improvement in one fifth of the 2024 time, sustained for several months |
| If crossed | Stronger safeguards and a documented risk case | At High, security controls to the High standard. At Critical, halt development until Critical safeguards exist |
Sources: the Mythos 5.2 system card's description of the RSP, the RSP page's version history, and the Preparedness Framework text. The RSP page summarizes the policy and does not reproduce its full text, so check the policy itself for exact definitions. OpenAI's framework may have newer versions than the one linked here.
What Anthropic says about its new models
The system card for Claude Mythos 5.2, dated November 9, is the most detailed public statement yet on where a lab sits against these lines.
Anthropic says Mythos 5.2 does not seem close to replacing its research staff, especially senior people. On the acceleration test, it fit a trend line to its own capability index, a fork of Epoch AI's Epoch Capabilities Index. The fitted change in slope is about 1.88 times (95% confidence interval 1.62 to 2.28), and 32 of 100 refits reach 2 times or more. Anthropic concludes the model does not double its internal rate of progress and so does not meet the RSP threshold, but calls that judgment uncertain and says it becomes harder with each advance. It plans to publish more on acceleration trends.
The card also says several of its internal measures of research acceleration are starting to show acceleration, and that METR, an outside evaluator, assessed it independently. METR's pre-release testing used a snapshot stronger on AI R&D than the released model, so its results may overstate what shipped. TechCrunch reports that a November 5 METR report put the speedup from internal models at roughly 2 times, with many uncertainties. We have not read that report.
Anthropic held back some AI research skills on purpose
The announcement says training included steps to limit research capabilities that could most directly accelerate AI development. In accelerator kernel optimization and LLM architecture improvement, Fable 5.2 performs at about the level of Opus 5.5. Agentic coding, hardware development and general AI and ML research, including interpretability, were not affected.
The system card adds a second layer. Safeguards on Fable 5.2 cover a narrow set of frontier-LLM development tasks, such as kernel development on certain ML accelerators. Blocked requests fall back to Opus 5. Anthropic says these should not affect most ordinary AI or ML work.
For earlier coverage, see our report on Anthropic's S-1 risk factor and our launch report on the new models.
What research shows about self-improving systems
Research on small, bounded loops is moving. A September 2026 paper, Recursive self-improvement of AI research agents by Srikanth and colleagues, describes AIDE², an agent that proposes edits to its own code, tests them on AI R&D tasks and keeps the best. In one autonomous 8-day run it found seven successive improvements. The gains carried over to four held-out benchmarks, where the best agent matched or beat a human-built production agent. The authors report reward hacking on a held-out task family fell from 55% to 32% during the run. The main result comes from a single run, which limits how far to generalize it.
Theory is older. A 2018 paper by Wenyi Wang gives a formal definition of one class of recursive self-improving systems and argues, with simulations, that efficient ones can exist. It is a formal argument for a restricted class, not a forecast about frontier models.
What to watch
Anthropic has said more is coming on acceleration trends and what they mean under its policy. The next system cards from Anthropic and OpenAI will show whether the estimates keep rising, and whether the labs move the line or cross it. The Mythos 5.2 card is one lab's self-assessment, with outside input from METR, and the labs set their own thresholds.
Frequently asked questions
What is recursive self-improvement in AI?
It is a loop in which an AI system helps build a better AI system, which then builds a better one, so progress compounds. OpenAI's framework defines the extreme case as fully automated AI R&D.
What does RSI stand for in AI?
RSI stands for recursive self-improvement. Anthropic's Mythos 5.2 system card uses the abbreviation when it discusses the risks of speeding up the overall pace of model development.
Is recursive self-improvement happening now?
Not in the full sense. Anthropic says over 90% of merged changes to its core codebase are written by Claude with human guidance, but it says Mythos 5.2 does not seem close to replacing its research staff and does not double its internal rate of progress.
Why is recursive self-improvement considered risky?
OpenAI's framework says a major acceleration in AI R&D could rapidly raise the rate at which new capabilities and risks emerge, beyond what current oversight can handle. Anthropic's card says it is concerned about accelerating the pace of model development.
Sources
- Claude Fable 5.2 and Mythos 5.2Anthropic
- Claude Fable 5.2 and Mythos 5.2 system cardAnthropic
- Anthropic Responsible Scaling PolicyAnthropic
- Preparedness Framework, version 2OpenAI
- Recursive self-improvement of AI research agentsarXiv
- A Formulation of Recursive Self-Improvement and Its Possible EfficiencyarXiv
- Anthropic releases new Mythos and Fable models amid RSI concernsTechCrunch

