The artificial intelligence industry is approaching a point where models do more than assist developers: they increasingly speed up the creation of successor systems. In a recent investor filing, Anthropic explicitly listed the risk of recursive self-improvement (RSI), and its internal measurements show that Claude already plays a substantial role in the company’s engineering workflow.
What Anthropic says in the filing
Anthropic states that its current models have not yet crossed the company’s internal safety threshold at which an AI would design and train new state‑of‑the‑art models without human oversight. Nevertheless, the filing makes clear that AI already accelerates several crucial steps of AI development (code writing, testing, debugging and trainer‑code optimization), and if that acceleration continues, development could outpace what human review can safely monitor.
Key performance and efficiency figures
-
Trainer‑code optimization benchmark: a skilled human AI researcher achieved roughly a fourfold speedup in 4–8 hours on a small model training script. In the same controlled task, Claude Opus 4 (May 2025) achieved about a threefold speedup, while Claude Mythos Preview (April 2026) reached a 52‑fold speedup. Anthropic notes this is a confined engineering test rather than full model training, but it highlights gains on the execution side.
-
Duration of autonomous work: successive Claude releases have handled progressively longer, multi‑step tasks. METR measurements reported that Claude Opus 3 (March 2024) reliably handled tasks comparable to about four minutes of human work; Claude Sonnet 3.7 one year later reached roughly 1.5‑hour tasks; Claude Opus 4.6 the following year reached 12‑hour workflows; and Mythos Preview entered at least the 16‑hour range.
-
Code contributions: Anthropic reports that in September 2026, more than 90 percent of accepted new code changes in its codebase during the period were written by Claude. Methodological note: this percentage refers to accepted new changes in the given period, not to the company’s entire multi‑year codebase.
-
Per‑engineer code output: the filing states that in Q3 2026 the daily amount of code per engineer was more than 45 times the 2024 average, reflecting the increased throughput enabled by Claude‑prepared solutions.
Practical implications
In practice, when an Anthropic engineer updates internal systems, the proposed solution increasingly comes from Claude; the human typically directs, reviews and approves the result. Anthropic emphasizes this does not mean Claude today autonomously invents, trains and deploys its own successor model without human intervention. Humans continue to set objectives, interpret outcomes and choose research directions.
Why this is a material investment risk
By listing RSI as a risk factor in its investor filing, Anthropic elevates a long‑standing technical and ethical debate into formal financial disclosure. The immediate concern is not acceptance or rejection of the most extreme superintelligence scenarios; it is the nearer‑term issue that development speed is rising in areas where safety checks, corporate accountability and regulation have traditionally assumed human timescales. If automation accelerates model development faster than oversight can adapt, governance and safety mechanisms may lag.
Conclusion
Anthropic’s filing is significant because it frames the growing role of AI in its own development as an investor‑level risk. Internal metrics show Claude already substantially increases engineering throughput and the scope of tasks handled autonomously, but the company maintains that current models have not reached a state of independent, unsupervised model creation. Still, the trends underline the need to consider how safety practices, corporate procedures and regulations should evolve as development accelerates.
