How Anthropic's RSP decides a capability threshold is met, and what follows

Anthropic's Amendment No. 1 to Form S-1, filed November 2, 2026, says Anthropic "may imminently meet our Responsible Scaling Policy’s definition of ‘dramatically accelerated progress attributable to AI.’"

our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.”

The amended S-1 says: "We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data."

We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data.

RSP v3.4 section 1 describes a three-column table: the left column identifies capability thresholds that would call for heightened mitigations; the middle column gives an overview of planned mitigations; the right column describes recommendations for industry-wide safety at each threshold. The changelog lists v3.4 as July 8, 2026.

We lay this out in a three-column table. The left column identifies capability thresholds that would call for heightened mitigations. The middle column provides an overview of our planned mitigations, which we have set forth more fully in our Frontier Safety Roadmap and other sections of this policy. The right column describes our recommendations for industry-wide safety at each threshold. ... July 8, 2026 (RSP v3.4)

The RSP table row reads: "Automated R&D in key domains. AI systems that can fully automate, or otherwise dramatically accelerate, the work of large, top-tier teams of human researchers in domains where fast progress could cause threats to international security and/or rapid disruptions to the global balance of power--for example, energy, robotics, weapons development, and AI itself."

Automated R&D in key domains. AI systems that can fully automate, or otherwise dramatically accelerate, the work of large, top-tier teams of human researchers in domains where fast progress could cause threats to international security and/or rapid disruptions to the global balance of power--for example, energy, robotics, weapons development, and AI itself.

The row continues: "For now, our evaluations will focus specifically on AI R&D, as this domain likely plays to AI systems' current strengths and is more tractable to assess than capabilities in other domains."

For now, our evaluations will focus specifically on AI R&D, as this domain likely plays to AI systems' current strengths and is more tractable to assess than capabilities in other domains.

RSP v3.4: "We will consider this threshold to be met if we determine that either (1) our models would be able to fully substitute for our entire set of Research Scientists and Research Engineers, at competitive costs (i.e., within a factor of 5); or (2) there is "dramatic acceleration" of the pace of AI progress for reasons that likely relate to the automation of AI R&D."

We will consider this threshold to be met if we determine that either (1) our models would be able to fully substitute for our entire set of Research Scientists and Research Engineers, at competitive costs (i.e., within a factor of 5); or (2) there is "dramatic acceleration" of the pace of AI progress for reasons that likely relate to the automation of AI R&D.

RSP v3.4 says scenario (2) has occurred where (a) Anthropic observes or expects double the rate of progress in aggregate capabilities compared to both the rate it would expect and the fastest rate of extended progress it has observed in the absence of significant AI contributions to AI R&D, and (b) it is plausible that this doubling is substantially attributable to the automation of research and/or engineering.

We would consider scenario (2) to have occurred where (a) we observe or expect double the rate of progress in AI aggregate capabilities compared to both the rate we'd expect and the fastest rate of extended progress we've observed in the absence of significant AI contributions to AI R&D and (b) it is plausible that this doubling is substantially attributable to the automation of research and/or engineering

RSP v3.4 footnote 8: "Double the rate of progress" means "as much progress in one year as one would see in two years at baseline"; it "is not the same idea as doubling researchers’ productivity".

"Double the rate of progress" means "as much progress in one year as one would see in two years at baseline." ... This is not the same idea as "doubling researchers' productivity," since doubling inputs does not necessarily double the rate of progress.

The RSP v3.4 changelog says the update revises the Automated R&D threshold so that it would not be considered crossed if the overall rate of AI progress were constant or slowing but estimated to be dramatically faster than without advanced AI tools, and that the threshold "is intended to capture the onset of dramatic recursive self-improvement, and has proven difficult to operationalize."

It revises the Automated R&D capability threshold in light of further issues that have come up in discussion - particularly the question of whether we'd consider the threshold to be crossed if the overall rate of AI progress were constant or slowing, but we estimated that it was dramatically faster than it would be in the absence of advanced AI tools. We would not ... This threshold is intended to capture the onset of dramatic recursive self-improvement, and has proven difficult to operationalize.

For the Automated R&D row, the industry-wide recommendation reads that a frontier developer should make a strong argument that no user or team of users (including those backed by top-tier states) will become significantly more likely to cause catastrophic harm via their usage of product surfaces or via theft of model weights, and that accomplishing this would likely mean security roughly in line with RAND SL4.

A frontier developer should make a strong argument that: No user or team of users (including those backed by top-tier states) will become significantly more likely to cause catastrophic harm via their usage of product surfaces or via theft of model weights. ... Accomplishing this would likely mean security roughly in line with RAND SL4.

RSP v3.4 section 1 says: "we cannot unilaterally and unconditionally commit to staying in line with the industry-wide recommendations in the right column."

In particular, we cannot unilaterally and unconditionally commit to staying in line with the industry-wide recommendations in the right column.

RSP v3.4 section 3.1 says: "We will publish a Risk Report every 3-6 months."

Timing. We will publish a Risk Report every 3-6 months.

RSP v3.4 section 3.1 says Anthropic will publish an analysis of an individual model's risks within 30 days of determining that it has an internally deployed model that could pose risks related to high-stakes misalignment or automated R&D that significantly exceed those of all models for which it has publicly analyzed such risks.

Within 30 days of determining that we have an internally deployed model that (1) could pose risks related to high-stakes misalignment or automated R&D that (2) significantly exceed those of all models for which we have publicly analyzed such risks

RSP v3.4: a model is "highly capable" if Anthropic concludes that it crosses the threshold for automated AI R&D described in Section 1; at a minimum it will complete a full external review with at least one external reviewer anytime a Risk Report covers highly capable models and is significantly redacted.

At a minimum, we will complete a full external review process (described below) with at least one external reviewer anytime a Risk Report covers highly capable models and is significantly redacted ... A model is "highly capable" if we conclude that it crosses the threshold for automated AI R&D described in Section 1.

RSP v3.4 Appendix A, scenario "Anthropic in the lead": having developed or imminently developing a highly capable model with clear evidence no other competitor will soon develop one, Anthropic commits: "We will require a strong argument that catastrophic risk is contained... We will delay AI development and deployment as needed to achieve this, until and unless we no longer believe we have a significant lead."

Anthropic in the lead. We have developed or will imminently develop a highly capable model; and we have clear evidence that no other competitor will soon develop such a model. ... We will require a strong argument that catastrophic risk is contained, along the lines of our recommendations for industry-wide safety (see Section 1). We will delay AI development and deployment as needed to achieve this, until and unless we no longer believe we have a significant lead.

RSP v3.4 introduction: "Our Responsible Scaling Policy (RSP) is our voluntary framework for managing catastrophic risks from advanced AI systems."

Our Responsible Scaling Policy (RSP) is our voluntary framework for managing catastrophic risks from advanced AI systems.

Anthropic's RSP page lists Version 3.4 as effective July 8, 2026, and describes the RSP as a framework Anthropic has refined since its first version in September 2023.