Anthropic has admitted something concerning: its AI development has accelerated before, and the company didn't recognize the change until after it happened.
Now the Claude developer thinks another acceleration could be close.
In a November 2 SEC filing, Anthropic warned that future AI models may soon reach its internal threshold for “dramatically accelerated progress attributable to AI.”
The company hasn't designated any current model as crossing that threshold. But its admission raises an obvious question: if Anthropic has already struggled to recognize changes in its development speed, what happens when AI starts taking on even more of the research?
Claude is already helping build the next Claude
Anthropic is increasingly using Claude to develop newer versions of itself. The model writes code, helps run experiments, and assists researchers with training more capable AI systems.
Here's how the process works.
An engineer might ask Claude to write software for a training experiment. Researchers review the results and use what they've learned to improve the next model. That newer version could then help with more complex experiments.
Repeat the process, and AI could eventually accelerate its own development.
Researchers call this recursive self-improvement (RSI). At its most advanced, an AI system could independently design successive generations of models, potentially speeding up development each time.
Anthropic says it hasn't reached full autonomous RSI, although Claude is already doing a substantial amount of engineering work.
According to Anthropic's research, Claude wrote more than 80% of the code merged into its production codebase as of May 2026.
By September, the company's latest SEC filing says Claude was responsible for more than 90% of code contributions. Average daily code contributions per engineer during the third quarter were also more than 45 times their 2024 level.
Those numbers sound impressive, although there's a limitation worth understanding.
Writing more code doesn't necessarily mean researchers are developing better AI at the same rate. Engineers still need to decide which experiments are worth pursuing and interpret the results.
And those decisions are considerably harder to automate.
So, how close is Anthropic to AI improving itself?
Anthropic has a specific way of measuring how much AI contributes to its development.
Under its Responsible Scaling Policy, the company considers its automated research threshold reached if AI can replace its entire research staff at a competitive cost or dramatically accelerate overall AI capabilities.
For the acceleration measure, progress would need to double relative to both the expected pace and the fastest sustained historical pace before substantial AI involvement. That historical comparison covers at least three model generations.
Think of it as achieving two years of expected AI capability improvements in one year.
Anthropic would also need evidence that automated research was substantially responsible for the increase, rather than additional computing power or more researchers.
The November filing suggests future models could meet this definition soon. However, Anthropic is still assessing its internal data, and its forecast doesn't establish that Claude can independently develop its successors.
Independent researchers are seeing progress, but also limitations
Before taking Anthropic's prediction at face value, it's worth looking at what independent researchers have found.
In September, AI evaluation organization METR tested Claude Opus 5.5 on difficult research tasks.
The model showed improvements in AI research capabilities, but METR considered it unlikely to automate the entire research process.
Claude still struggled with work that requires experienced researchers to judge promising ideas and decide what to investigate next.
A separate preliminary METR assessment estimated about 1.5 times acceleration in overall AI capability progress attributable to AI, although the researchers acknowledged substantial uncertainty.
The assessment didn't specify the period covered by that estimate, and the full supporting evidence wasn't available to the team evaluating Opus 5.5.
These results concern an earlier model, so they can't establish what Anthropic's future systems will achieve.
Still, they offer a more cautious picture of how much research AI can currently take on.
What happens when safety research can't keep up?
Anthropic's biggest concern is that models could improve faster than researchers can understand their capabilities.
Its SEC filing acknowledges that some measurements reveal acceleration only after it has occurred, making it difficult to anticipate when additional safeguards are needed.
Consider an AI system that helps develop training software for the next Claude model. If its work contains an error that researchers overlook, the mistake could affect the new model.
The concern becomes greater if the same AI systems also help conduct safety evaluations.
Anthropic says it doesn't currently expect to safely allow Claude to build successive frontier models entirely autonomously. The company is considering policy updates to preserve meaningful human supervision.
There's a business incentive to move quickly, too.
If AI-assisted research helps Anthropic develop better models faster, rival labs may face pressure to accelerate their own work. But each new generation could also require more extensive testing, potentially affecting release schedules.
Anthropic acknowledges this trade-off in its filing, including the possibility that safety precautions could slow its progress relative to competitors.
What would prove AI is accelerating its own development?
The next step is figuring out how much of AI's progress actually comes from AI itself.
In an October proposal, METR outlined how independent assessors could investigate these claims. Researchers would need access to unreleased models and historical capability measurements to determine whether development is accelerating.
They would also need to separate improvements attributable to AI automation from gains caused by additional computing resources or human research.
And there's another question worth asking: would their evaluation methods detect a much larger acceleration before it became apparent through newer model releases?
Anthropic has evidence that AI is already helping researchers develop newer models faster. Whether this eventually becomes autonomous recursive self-improvement remains uncertain.
For now, the question is whether researchers can measure AI's growing contribution to development accurately enough to identify new capabilities and evaluate their risks before the next generation arrives.
