Skip to main content
v2026.11,986 entries · CC-BY 4.0
CASRAIAsk CASRAI · AI assistant for researchAsk a research question. Answers come with their sources.Funder rules, compliance, methods and publishing. Answers link the sources they rest on.
NIKOLAI elementN3 · Thresholds and checkpointsProposednikolai-v0.3

Capability Threshold

Definition

A capability threshold is a pre-defined level of AI model capability, such as being able to meaningfully help someone build a biological weapon, at which a developer commits to stronger safeguards or other decisions before further development or deployment. It describes what a model can do, not how much risk remains after safeguards.

Also called: Critical Capability Level (CCL), Critical Capability Threshold, Enabling capability threshold, Frontier Capability Threshold, High / Critical capability threshold, Capability or usage threshold

NIKOLAI's proposed operational definition

NIKOLAI editorial proposal (unsourced): a capability-threshold record states a level of model capability (or capability plus usage) at which specified additional safeguards or decisions become required, together with its disclosure status (quantified, qualitative, referenced-but-undefined, or classified).

The operational definition is CASRAI's own proposal, not a definition any named organisation has agreed to. See what NIKOLAI is and is not, or browse the AI safety glossary.

Examples

Examples from published documents

  • Anthropic's novel chemical and biological weapons threshold covers AI systems that can functionally substitute for the scarce human expertise that is currently the primary barrier to developing such weapons (Anthropic RSP v3.4, section 1)
  • OpenAI's High threshold for AI Self-improvement is a model whose impact equals giving every OpenAI researcher a highly performant mid-career research engineer assistant, relative to a 2024 baseline (OpenAI Preparedness Framework v2, Table 1)
  • Magic set its threshold at models exceeding 50% accuracy on LiveCodeBench at the end of a training run (Magic AGI Readiness Policy, Threshold Definition)

Counter-examples

Not to be confused with

  • A risk threshold or overall risk level (for example Meta's "High" or "Critical" risk thresholds) measures the risk a deployment poses, often after safeguards; a capability threshold measures what the model can do before safeguards.
  • A compute threshold, such as a training-compute figure that decides whether a law applies to a model at all, is a coverage test (see coverage-scope-threshold), not a capability threshold.

Related terms

Related AI safety terms

Terms defined under this element that are not NIKOLAI elements themselves, each as the cited document uses it. Every term is also listed in the AI safety glossary.

AI Safety Level (ASL)
AI Safety Level (ASL) is Anthropic's label for graded sets of safety and security safeguards, such as ASL-3. Earlier versions of its Responsible Scaling Policy also used ASLs to classify models; version 3.4 uses ASLs only for present safeguard levels.
Source: Anthropic Responsible Scaling Policy v3.4
Critical Capability Level (CCL)
A Critical Capability Level (CCL) is Google DeepMind's term for a capability threshold: a capability level at which, absent mitigation measures, a frontier AI model may pose heightened risk of severe harm.
Source: Google DeepMind Frontier Safety Framework v3.1
Tracked Capability Level (TCL)
A Tracked Capability Level (TCL) is Google DeepMind's term for a capability threshold below a Critical Capability Level, marking where a model may pose significant but not severe harm. Mitigation and risk acceptance apply proportionately at a TCL.
Source: Google DeepMind Frontier Safety Framework v3.1
High capability
High capability is the lower of the two capability thresholds in OpenAI's Preparedness Framework: capabilities that significantly increase existing risk vectors for severe harm. Models that cross it need safeguards that sufficiently minimise the risk before deployment.
Source: OpenAI Preparedness Framework v2
Critical capability
Critical capability is the upper capability threshold in OpenAI's Preparedness Framework: capabilities that present a meaningful risk of a qualitatively new threat vector for severe harm with no ready precedent. It requires safeguards even during development.
Source: OpenAI Preparedness Framework v2
Tracked Category
A Tracked Category is a capability area that OpenAI's Preparedness Framework measures for every covered deployment and prepares safeguards for. The current ones are Biological and Chemical, Cybersecurity, and AI Self-improvement; less mature areas are Research Categories.
Source: OpenAI Preparedness Framework v2
Intolerable risk
Intolerable risk is the level of severe risk an organisation will not accept unless it is adequately mitigated. Signatories of the 2024 Seoul Frontier AI Safety Commitments agreed to set thresholds marking intolerable risk and to monitor how close models come to them.
Source: Frontier AI Safety Commitments (AI Seoul Summit 2024)

Source of record

Where this definition comes from

Crosswalk

How named organisations use this concept

Every row below is a shadow mapping. A shadow row is CASRAI's own reading of a published document. No lab, evaluator or regulator named on a shadow row has declared, endorsed, or been consulted on it. That changes only when an organisation files its own Mapping Declaration.
OrganisationTheir term, as publishedMatch & verificationSource
AnthropicShadow mapping
Anthropic Risk Report, August 2026 / RSP v3.4
“"AI R&D threshold (current)": met if "(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" (§1.3.1, §3.2). CB-2 threshold: "AI systems that can functionally substitute for the scarce human expertise that is currently the primary barrier to novel development of chemical and biological weapons with potential for catastrophic harm ..." (§1.3.2). RSP §1 table header: "Capability or usage threshold."”
The threshold wording appears in both RSP v3.4 (section 1 table, whose first column is headed "Capability or usage threshold") and the Risk Report, which applies it.
exactEQ
confidence: high
Anthropic Risk Report, August 2026
OpenAIShadow mapping
Preparedness Framework v2 / Frontier Governance Framework
“"Capability thresholds: concretely describe things an AI system might be able to help someone do or might be able to do on its own that could meaningfully increase risk of severe harm." (§2.2). FGF: the term capability threshold is used but not formally defined; "Risk tiers" are the published scale (§2.2, §2.3).”exactEQ
confidence: high
OpenAI Preparedness Framework v2
Google DeepMindShadow mapping
Frontier Safety Framework v3.1
“"Critical Capability Levels (CCLs): are the main capability thresholds around which we have built the Framework process." "Tracked Capability Levels (TCLs): are capability thresholds which capture a lower level of risks than our CCLs." (glossary, p.18)”exactEQ
confidence: high
Google DeepMind Frontier Safety Framework v3.1
xAIShadow mapping
xAI Frontier AI Framework (30 Jun 2026) / Grok 4.6 card / earlier xAI frameworks
“June 2026 framework: "When analyzing whether a threshold has been reached ..." (s.2.3), with no thresholds stated. Grok 4.6 card: "FAIF safety thresholds" cited with no values given. The earlier Dec 2025 FAIF and Aug 2025 RMF used deployment acceptance criteria instead: "maintaining an answer rate of less than 1 out of 20 on restricted queries" and "a dishonesty rate of less than 1 out of 2 on MASK."”
PDF metadata /Title of the 30 Jun 2026 framework reads "Privileged/Confidential DRAFT working FRAMEWORK DOC"; no xAI statement disambiguating draft vs. final was found.
noneDU
confidence: medium
xAI Frontier AI Framework (30 Jun 2026, draft-marked)
EUShadow mapping
EU GPAI Code of Practice, Safety and Security chapter
“Measure 4.1: for each identified systemic risk, Signatories "define appropriate systemic risk tiers that: (i) are defined in terms of model capabilities, and may additionally incorporate model propensities, risk estimates, and/or other suitable metrics; (ii) are measurable; and (iii) comprise at least one systemic risk tier that has not been reached by the model." Content of tiers left to each Signatory; no shared tier scale published in the chapter.”exactEQ
confidence: high
EU GPAI Code of Practice, Safety and Security chapter
California SB 53Shadow mapping
California SB 53
“Frameworks must describe "(2) Defining and assessing thresholds used by the large frontier developer to identify and assess whether a frontier model has capabilities that could pose a catastrophic risk, which may include multiple-tiered thresholds." (22757.12(a)). "Threshold" itself is not defined by the statute.”noneDU
confidence: medium
California SB 53
US Government (Executive Order 14409)Shadow mapping
EO 14409 / Seoul Frontier AI Safety Commitments
“EO 14409 designation threshold is classified (cyber domain). Seoul Commitment II: thresholds are "at which severe risks posed by a model or system, unless adequately mitigated, would be deemed intolerable" — a narrower, outcome-gated framing than a capability threshold proper.”narrowNR
confidence: medium
Executive Order 14409
METRShadow mapping
METR Common Elements
“"Capability Thresholds": "Thresholds at which specific AI capabilities would pose severe risk and require new mitigations."”exactEQ
confidence: high
METR (common-elements)
Frontier Model ForumShadow mapping
FMF Risk Taxonomy and Thresholds
“"Enabling Capability Thresholds (also called 'critical capability levels' or 'capability thresholds'): Abilities that could potentially enable extreme harms if the model is deployed without additional safeguards." (s3.1, p.10)”exactEQ
confidence: high
FMF Risk Taxonomy and Thresholds
AmazonShadow mapping
Amazon Frontier Model Safety Framework
“"Critical Capability Thresholds": "a set of model capabilities that have the potential to cause significant harm to the public if misused" (Overview, p.1); one qualitative threshold per domain (CBRN, Offensive Cyber Operations, Automated AI R&D), stated as an uplift-based description rather than a benchmark score.”exactEQ
confidence: high
Amazon Frontier Model Safety Framework
MagicShadow mapping
Magic AGI Readiness Policy
“"when, at the end of a training run, our models exceed a threshold of 50% accuracy on LiveCodeBench, we will trigger our commitment to incorporate a full system of dangerous capabilities evaluations and planned mitigations into our AGI Readiness Policy" (Threshold Definition), with the contemporaneous public-model baseline "Claude-3.5-Sonnet 48.8%" given for comparison.”
The page carries the banner "Note: This policy is outdated. We're working on an update."
exactEQ
confidence: high
Magic AGI Readiness Policy
What do these codes mean?
exact
The source term is equivalent to this element
close
The source term is close but not equivalent to this element
broad
The source term is broader than this element
narrow
The source term is narrower than this element
none
No mapping claim — used for false-friend and declared-but-undefined rows
EQ
Equivalent
CL
Close
BR
Source is broader than the element
NR
Source is narrower than the element
FF
False friend — same or similar label, different meaning
DU
Declared but undefined by the source
UV
Unverified

Related, not mapped

Pointers that are not crosswalk claims

These sources mention this concept but do not define or map it clearly enough to count as a crosswalk row — noted here so the research is visible without overstating it as a mapping.

  • Meta

    "Risk thresholds: are the incremental levels of risk that a Frontier AI model might pose towards realization of a catastrophic outcome" (Appendix I) — these are outcome-based risk levels, not capability thresholds; capability enters only through separately-named "Enabling capabilities" and "Operational threshold" (cyber).

    Meta Advanced AI Scaling Framework v2

Gap

SB 53 requires thresholds but defines none; xAI's current framework and the Grok 4.6 card cite thresholds that are not published; Amazon's thresholds are qualitative only, with undefined terms such as "material uplift" and "reliably." Numeric public values are rare: Magic discloses a benchmark threshold (50% on LiveCodeBench), Meta discloses evaluation cut-offs such as "< 75% pass@10" on simple capture-the-flag challenges, and Google DeepMind's Gemini 3.7 Flash report discloses a 90% safety margin for its ML R&D automation level. NIKOLAI proposes a `threshold.disclosureStatus` value set (quantified; qualitative; referenced-but-undefined; classified).

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →