Anthropic Warns It May Soon Hit Its Own Recursive Self-Improvement Threshold

Anthropic told IPO investors on November 2 that it may imminently meet its own threshold for dramatically accelerated, AI-driven progress — the point at which models deliver roughly two years of baseline advancement in a single year. The disclosure sits in an amended prospectus from a company that has spent nearly $28 billion on research and development since the start of 2024, and it arrives alongside an unusual admission: the CEO wants the pace to slow.

The spending curve is steep. Anthropic laid out $2.3 billion on R&D in 2024, $6.3 billion in 2025 and $19.2 billion in the first nine months of 2026 — with about $9.5 billion of the 2026 figure landing in the third quarter alone. For frame of reference, the nine-month 2026 outlay is roughly triple the full-year 2025 total.

What is 'closing the loop'?

Anthropic already uses Claude, its flagship model, to run experiments, prepare training data, write and review code, automate evaluations and safety audits, and test successor models. The company brands this internal reuse of its own AI as "closing the loop." Measured at list prices, the value of Anthropic's internal model use grew more than 250-fold between January and September 2026, according to the filing.

That 250-fold figure deserves scrutiny. It is a list-price valuation of internal consumption, not audited productivity data, and it starts from what was presumably a small January base. The prospectus itself concedes the company is still assessing whether any model has actually crossed the Responsible Scaling Policy threshold; no current model has been designated as having done so.

How close is recursive self-improvement?

External observers offer diverging timelines. Rayan Krishnan, co-founder and CEO of evaluation firm Vals AI, predicted in an October 3 interview with MTS that Anthropic could reach full recursive self-improvement around July or August 2027, based on tasks in his company's RSI Index. "It means that the work required by machine learning engineers at Anthropic would be automated," he said — though he acknowledged that chip design and other constraints could limit further acceleration.

John Schulman, chief scientist at Thinking Machines, flagged a more immediate bottleneck in a September 11 Dwarkesh Podcast discussion: models still struggle to judge and verify their own output. "Even if the model can write way more code than a person, it doesn't make you 100X more productive," he said.

Inside the frontier labs, estimates are converging on the near term, with caveats attached to nearly every one:

• Anthropic co-founder Jack Clark wrote in mid-October that he now considers full RSI more likely in 2027 than 2028, revising an earlier estimate.
• Alignment Science Lead Evan Hubinger has said the process is "happening faster than we thought."
• Most of roughly half-dozen researchers at top AI companies interviewed for a Gladstone AI report in October expect RSI to begin within 3 to 12 months, "and possibly sooner, assuming there's no U.S. government intervention or slowdown."
• Google CEO Sundar Pichai said in October that Google runs "RSI loops along with AI researchers in various aspects of model development," while remaining "pretty far from the end state of RSI."

Why is the CEO calling for brakes?

In his September essay "We Must Pace the Frontier," Anthropic CEO Dario Amodei wrote: "We must slow the pace at which we improve the capabilities of AI models." He argued that safeguards need time to catch up with AI-assisted development, and he proposed outside evaluators with continuing access to frontier labs plus coordinated limits on capability advances — potentially including caps on companies' internal use of AI to improve AI.

The prospectus quantifies the competitive cost of that restraint plainly: "These measures could slow our capability gains relative to competitors that do not adopt similar constraints." It also warns that AI-assisted research could advance model capabilities faster than Anthropic can develop and validate safeguards.

Anthropic has committed to embedded external evaluators, but the broader pacing proposals require buy-in from rival companies and governments — commitments that do not yet exist. The company expects human supervision to remain necessary even if Claude becomes technically capable of building successive frontier models independently, and it is considering policy changes to keep people in control of that development. Whether self-imposed brakes survive contact with a $19-billion-a-quarter R&D race is the question investors in the IPO will effectively be underwriting.