Anthropic safety capability thresholds
Loading…
This market contains two independent events, each resolving to the earliest month in which the condition is met:
- AI R&D-4: Anthropic publicly reports (in an RSP update, Risk Report, System Card, or blog post) that one of its models meets or exceeds the AI R&D-4 capability threshold as defined in the Responsible Scaling Policy. Under RSP v2.1, this is ‘the ability to fully automate the work of an entry-level, remote-only Researcher at Anthropic.’ Under RSP v3.0, the equivalent is a model that could ‘compress two years of 2018–2024 AI progress into a single year.’
- CBRN-4: Anthropic publicly reports that one of its models meets or exceeds the CBRN-4 capability threshold. Under RSP v2.1, this is ‘the ability to substantially uplift CBRN development capabilities of moderately resourced state programs (with relevant expert teams), such as by novel weapons design, substantially accelerating existing processes, or dramatic reduction in technical barriers.’
Each event resolves independently. Resolution requires an official Anthropic publication (not leaked documents or third-party evals). If Anthropic discontinues the RSP or renames capability levels, a substantially equivalent assessment under the successor framework counts.
Bars include events up to their close, 23:59:59 UTC: “By end of Jun 2025” includes everything up to 30 Jun 2025, 23:59:59 UTC.
Anthropic judged its newest model, Claude Mythos 5.2, just short of its automated AI R&D threshold: AI-driven acceleration of about 1.9 times against a bar of 2. In August it said the line may be crossed within a year. The novel-bioweapon threshold is not met, though Anthropic called its June verdict a close one.
Reference classes
- Claude Mythos 5.2 AI-driven acceleration estimate (9 Nov 2026): about 1.88x (95% CI 1.62x to 2.28x) against a 2x threshold; 32 of 100 refits reach 2x
- August 2026 Risk Report on automated AI R&D: not yet 2x faster; “we may cross this threshold in the coming year”
- Claude Mythos 5 on CB-2, the successor to CBRN-4 (Jun 2026): not crossed, but “a much less clear and obvious judgment than with previous models”
- Anthropic system cards since Feb 2026 reporting either threshold crossed: 0 of 11
- RSP rewrites in 2026: v3.0 to v3.4 (Feb to Jul); AI R&D threshold reworded twice, CBRN-4 replaced by CB-2
- METR 50% task horizon, Claude Mythos Preview (Apr 2026): about 17 hours; doubling about every 4 months since 2023; suite unreliable above 16 hours
- Hazard of the framework becoming unresolvable: ~1.3%/year (unchanged)
Method
The starting probabilities of this market were calculated using logistic CDF models fitted to anchor points calibrated via Anthropic’s own RSP assessment history, cross-lab safety-framework patterns, and METR capability-trend data. A structural risk discount is applied to account for institutional uncertainty over longer horizons.
Events
This market tracks two milestones from Anthropic’s Responsible Scaling Policy (RSP):
- AI R&D-4: Anthropic publicly reports that a model meets the AI R&D-4 capability threshold. Under RSP v2.1 (Mar 2025), this is defined as “the ability to fully automate the work of an entry-level, remote-only Researcher at Anthropic.” Under RSP v3.0 (Feb 2026), the equivalent threshold is a model that could “compress two years of 2018–2024 AI progress into a single year” (~1000x effective compute scaleup). As of February 2026, Anthropic explicitly states that Claude Opus 4.6 does not cross this threshold, though ruling it out is getting harder.
- CBRN-4: Anthropic publicly reports that a model meets the CBRN-4 capability threshold. Under RSP v2.1, this is “the ability to substantially uplift CBRN development capabilities of moderately resourced state programs (with relevant expert teams), such as by novel weapons design, substantially accelerating existing processes, or dramatic reduction in technical barriers.” Under RSP v3.0, the equivalent is a model that can “significantly help threat actors (e.g., moderately resourced expert-backed teams) create/obtain and deploy chemical and/or biological weapons with potential for catastrophic damages far beyond those of past catastrophes such as COVID-19.” This is much harder than the CBRN-3 threshold (which targets individuals with basic STEM backgrounds and triggered ASL-3 protections in May 2025). ASL-4 safeguards have not yet been defined.
Anthropic RSP assessment history
Anthropic has been unusually transparent about threshold assessments, publishing regular RSP updates and risk reports. This timeline directly informs the calibration:
Key implications: The Feb 2026 assessment directly constrains near-term AI R&D-4 probability — the latest flagship model was explicitly assessed and found to fall short. The next realistic opportunity is a future model generation, likely late 2026 at earliest. For CBRN-4, the gap between CBRN-3 (triggered May 2025) and CBRN-4 is very large: CBRN-3 targets non-experts while CBRN-4 requires uplifting state-backed expert teams.
METR capability trend
METR’s March 2025 analysis found that frontier-agent task horizon has been doubling roughly every 7 months. Current systems achieve ~1-hour task horizon at 50% reliability, with <10% success on tasks over ~4 hours. The AI R&D-4 threshold (“fully automate entry-level researcher”) likely requires reliable multi-day autonomous performance, suggesting this crossing is still some years away even with steep improvement trends.
Cross-lab reference points
- OpenAI: “High” AI R&D threshold (mid-career ML research engineer) not yet reached. o3 reportedly “medium.” GPT-4o rated “low” in bio/cyber.
- Google DeepMind: No CCL-1 threshold crossed through Gemini 3.1 Pro (Feb 2026). CBRN and Cyber alert thresholds reached but CCLs confirmed not met.
The cross-lab pattern is consistent: intermediate capability thresholds are being approached or crossed, but the most severe levels remain unmet at all labs.
Methodology
For each event, we set anchor points (month, cumulative probability) balancing:
- Anthropic’s own assessments: the Feb 2026 “does not cross AI R&D-4” statement and the May 2025 “ASL-4 ruled out” for CBRN directly constrain near-term probabilities
- METR task-horizon trend (~7-month doubling)
- Cross-lab threshold history: no lab has publicly reported a top-level crossing
- RSP framework churn: three major rewrites in 2.5 years increases structural uncertainty
A logistic CDF is fitted to each event’s anchors via least-squares. A structural / institutional risk discount is then applied multiplicatively: ~1.3%/year constant hazard rate for “framework becomes unresolvable” (RSP restructuring, threshold redefinition, reporting changes), calibrated so cumulative structural risk reaches ~7% at the 5.5-year horizon.
Key assumptions
- Anthropic continues publishing capability assessments (RSP updates, Risk Reports, or equivalent)
- The RSP threshold concepts remain substantially similar even as specific definitions evolve (the structural discount accounts for risk they do not)
- AI R&D-4 progresses faster than CBRN-4 (stronger commercial incentive, more training signal, and closer to current capability levels)
- CBRN-4 requires a much larger capability jump from CBRN-3 than most other threshold gaps
Sources
- Anthropic RSP updates page
- RSP v2.1 (Mar 2025)
- RSP v3.0 (Feb 2026)
- ASL-3 activation announcement (May 2025)
- METR: Measuring AI Ability to Complete Long Tasks (Mar 2025)
Liquidity Adjustment Log
- scheduled | at: 2026-09-15T14:21:52Z | effective: 2026-09-15T14:31:38Z | reduction: 50% | before: 23013.2428 | after: 11506.6216 | actor: nonstationary
- applied | at: 2026-09-15T14:21:52Z | effective: 2026-09-15T14:21:52Z | reduction: 50% | before: n/a | after: n/a | actor: nonstationary
Refitted on 17 November 2026 with the same method: judgement anchors, a logistic CDF fitted by least squares, and the ~1.3%/year framework discount. The curve is now conditional on no crossing by today. The assessment history was brought up to date: eleven system cards from Claude Sonnet 4.6 (Feb 2026) to Claude Mythos 5.2 (Nov 2026), the February and August 2026 Risk Reports, and RSP versions 3.1 to 3.4. None reports either threshold crossed. RSP v3.1 (Apr 2026) and v3.4 (Jul 2026) reworded the automated R&D threshold: full substitution for Anthropic's research staff, or AI-driven doubling of the rate of progress. RSP v3.3 (May 2026) replaced the CBRN-4 equivalent with CB-2, about novel chemical and biological weapons. These are treated as the successor thresholds.
The AI R&D anchors moved earlier. The Mythos 5.2 system card puts acceleration at about 1.88x against a 2x bar, and the August 2026 Risk Report says the threshold may be crossed within a year. The CBRN anchors moved much earlier. Anthropic called its June 2026 CB-2 verdict on Claude Mythos 5 “much less clear” than before, and OpenAI says it is preparing for its own Critical bio/chem threshold. METR's March 2025 figures were replaced. Its current data put Claude Mythos Preview at about 17 hours at 50% reliability, with horizons doubling about every 4 months since 2023 (METR).
Anthropic RSP assessment history
| Date | Event |
|---|---|
| Sep 2023 | RSP launched, defining ASL-1 to ASL-3 (ASL-4 and higher named but not yet defined). All current models, including Claude, assessed as ASL-2. |
| Oct 2024 | Major RSP update (v2). All models still ASL-2. |
| Mar 2025 | RSP v2.1: AI R&D thresholds disaggregated into AI R&D-4 and AI R&D-5; CBRN threshold detail added. |
| May 2025 | ASL-3 activated for Claude Opus 4, driven by CBRN-3 concerns (steadily increasing performance on Virology Capabilities Test, stronger virus acquisition task performance). Anthropic could not clearly rule out ASL-3 risks. ASL-4 explicitly ruled out for Opus 4. |
| Feb 2026 | Claude Opus 4.6 assessed (10 Feb): does not cross AI R&D-4, but “confidently ruling out this threshold is becoming increasingly difficult.” Sabotage risk report published for Opus 4.6 (10 Feb). RSP v3.0 (effective 24 Feb) published as a comprehensive rewrite: adds a Frontier Safety Roadmap and regular Risk Reports, and separates Anthropic's own commitments from an industry-wide capabilities-to-mitigations map. |
Checked against sources, November 2026. 2 rows corrected.
Sources
Liquidity over time
i Shows the market’s available liquidity over time. Liquidity can decay continuously and may change discretely if liquidity adjustments are applied.2 comments
Log in to comment.
Anthropic's Mythos 5.2 system card (November 9) says the model does not meet the automated AI R&D threshold, "though this assessment is uncertain": the fitted speed-up is 1.88x (95% CI 1.62–2.28x). CB-2 is not crossed either. Anthropic plans to "publish more soon" on what acceleration trends imply under the RSP. System card
Anthropic also says it limited, during training, the research capabilities that most directly accelerate AI development: on accelerator kernel optimisation and LLM architecture work, Fable 5.2 and Mythos 5.2 perform at about Opus 5.5's level. The card also reports trends from an unreleased, somewhat more capable model. Anthropic