TAI BUI
Back to curriculum
LearnPython

Anthropic Responsible Scaling Policy v3.0

RSP v3.0 went into effect February 24, 2026 and replaced the 2023 policy. It splits mitigations into two tiers: what Anthropic will do unilaterally, and what it frames as an industry-wide recommendation (including RAND SL-4 security standards). Frontier Safety Roadmaps and Risk Reports become standing documents instead of one-off deliverables, and the 2023 pause commitment is gone. The new AI R&D-4 threshold requires Anthropic to publish an affirmative case identifying misalignment risks and mitigations once a model crosses it. Claude Opus 4.6 does not cross it, but Anthropic states in the v3.0 announcement that "confidently ruling this out is becoming difficult." SaferAI rated the 2023 RSP at 2.2 and downgraded v3.0 to 1.9, which puts Anthropic in the "weak" category alongside OpenAI and DeepMind. Qualitative thresholds replaced the 2023 quantitative commitments, and removing the pause clause is the sharpest regression.

TIME: ~45 minutesPREREQUISITES: Phase 15 · 06 (AAR), Phase 15 · 07 (RSI)

Learning Objectives

  • Explain the two-tier mitigation schedule and tell an Anthropic commitment apart from an industry recommendation.
  • Define the AI R&D-4 threshold and state what Anthropic must publish once a model crosses it.
  • Describe what RSP v3.0 added (Frontier Safety Roadmaps, Risk Reports) and removed (the 2023 pause commitment).
  • Summarize SaferAI's downgrade from 2.2 to 1.9 and the four factors behind it.
  • Run main.py and read its any-two-of-three trigger rule as a reading aid, not a policy tool.

The Problem

Frontier labs publish scaling policies that are part technical document, part governance document, and part signal to regulators. RSP v3.0 is the current Anthropic one. It is worth reading closely, and not because it binds anyone (it doesn't). The framing shows how a lab thinks about catastrophic risk and how it explains trade-offs to the public.

The useful unit is the diff between v3.0 and the 2023 policy (v2.0). Added: Frontier Safety Roadmaps, Risk Reports, and the AI R&D-4 threshold. Removed: the 2023 pause commitment. Reframed: a two-tier mitigation schedule that separates Anthropic's unilateral actions from its industry recommendations. SaferAI, an outside reviewer, downgraded the score from 2.2 (the 2023 policy) to 1.9 (v3.0). A scaling policy can get less rigorous while looking more polished, and this diff shows how.

Warm-up0/3
1. In the Automated Alignment Research setup, why do the agents' logs live in a shared forum outside every sandbox?Warm-up
Three dashed sandboxes, each holding one agent, with arrows from every agent to a single shared forum box that sits outside all the sandboxes.
Typed schematic of the setup in the stem.
2. In the recursive self-improvement gap M(t) = C(t) - A(t), what does a growing M(t) indicate?Warm-up
A line chart of M(t) against time t that rises with a steepening slope, headed by the formula M(t) = C(t) - A(t).
Illustrative shape, no units.
3. A company publishes a voluntary safety policy that no regulator enforces. What is the practical consequence?Warm-up
A company box with an arrow to its published voluntary safety policy, which points to a public readers box; a regulator box below has a dashed line to the company marked with a question mark, above a dashed slot for the practical consequence.
Typed schematic of the stem's situation.
Check: The Problem0/1
1. What is the Responsible Scaling Policy (RSP)?Define
Two document boxes, the 2023 RSP called v2.0 and RSP v3.0 with 32 pages, joined by an arrow, above a dashed slot asking what the RSP is.
Document names and page count come from the lesson text.

The Concept

The two-tier mitigation schedule

  • Anthropic unilateral actions: what Anthropic will do regardless of what other labs do. Publishing an affirmative case before scaling past AI R&D-4, specific security measures, specific deployment gates.
  • Industry-wide recommendations: what Anthropic thinks the industry should do collectively. Includes RAND SL-4 security standards. These are not commitments on Anthropic's side; they are policy advocacy.

The two-tier structure was not in v2, so you have to check which column each commitment lives in. A security measure in the "industry-wide recommendation" column is not Anthropic's promise. It is Anthropic's hope.

The AI R&D-4 threshold

RSP v3.0 names this as the next threshold that matters: a model that could automate a substantial fraction of AI research at competitive cost. Once Anthropic believes a model crosses it, they must publish an affirmative case identifying misalignment risks and mitigations before continued scaling.

Claude Opus 4.6 does not cross it per the v3.0 announcement. The document adds: "confidently ruling this out is becoming difficult." That phrasing matters. It concedes that the threshold is close enough to be a live concern, not a speculative limit.

Lesson 6 (Automated Alignment Research) and Lesson 7 (Recursive Self-Improvement) feed directly into this threshold. Automated alignment researchers clearing research-quality bars is evidence that AI R&D-4 is approaching.

Frontier Safety Roadmaps and Risk Reports

v3.0 elevates two artifact types to standing documents:

  • Frontier Safety Roadmap: forward-looking document describing planned safety work, capability expectations, and mitigation research.
  • Risk Report: retrospective document on specific models after release, describing observed capability and residual risk.

Both are public and updated on a declared cadence. That lets you compare what Anthropic said it would do in a Roadmap against what it reports in a Risk Report.

Removing the pause clause

The 2023 RSP included an explicit pause commitment: if a model crossed specific capability thresholds, training would pause until mitigations were in place. v3.0 replaces the explicit pause with a softer formulation (publish an affirmative case, proceed if mitigations are adequate). SaferAI and other analysts called this out directly as the sharpest regression in the new document.

The policy argument for the change: quantitative thresholds in 2023 turned out to be unreachable by 2026-era capability benchmarks because the benchmarks themselves were re-scaled. The counter-argument: a pause clause in a scaling policy is a commitment device; removing it removes the credibility of the policy.

SaferAI's downgrade

SaferAI is an independent organization that rates RSP-style documents. Their public rating: 2023 Anthropic RSP scored 2.2 (out of a scale where 4.0 is the best current RSP and 1.0 is nominal). v3.0 scored 1.9. That moved Anthropic from "moderate" to "weak," alongside OpenAI and DeepMind.

The downgrade factors per SaferAI:

  • Qualitative thresholds replaced quantitative ones.
  • Pause commitment removed.
  • AI R&D-4 threshold mitigations are described as "affirmative case" rather than specific measures.
  • Review mechanisms depend on Anthropic's Safety Advisory Group, with limited independent oversight.

What this lesson is not

Compliance is not the point. RSP v3.0 is not a regulation, and nothing forces Anthropic to follow it. What matters is reading it with specificity and skepticism. Scaling policies are the main public signal frontier labs emit about their catastrophic-risk posture, and reading them well is a practical skill if your work depends on frontier capabilities.

Check: The Concept0/22
1. Which description matches the AI R&D-4 threshold?Define
A capability axis with a tick named AI R&D-4 and a dashed slot with a question mark for what a model past the tick can do.
Typed schematic.
2. What is an affirmative case in RSP v3.0?Define
A short list of three numbered section headings from main.py's affirmative-case template (capability inventory, evaluation-context gap, residual risk acknowledgement) followed by a note that further sections are not shown.
Three headings copied from the lesson's illustrative main.py, not from the real policy.
3. What is a Frontier Safety Roadmap?Define
RSP v3.0 linked by an arrow to a stack of cards titled Frontier Safety Roadmap, with a dashed slot asking what the document covers.
Typed schematic of the link named in the lesson's reading list.
4. What is a Risk Report?Define
RSP v3.0 linked by an arrow to a stack of cards titled Risk Report, with a dashed slot asking what the document covers.
Typed schematic of the link named in the lesson's reading list.
5. What does the two-tier mitigation schedule separate?Define
Two side-by-side panels named Tier 1 and Tier 2; Tier 1 lists specific security measures and specific deployment gates, Tier 2 lists RAND SL-4 security standards, with a question slot for what separates them.
Example items from the lesson.
6. What was the 2023 pause commitment?Define
A 2023 RSP document box, called v2.0, with an arrow to a box named 2023 pause commitment, above a dashed slot asking what that commitment promised.
Typed schematic.
7. What is the SaferAI rating of a scaling policy?Define
A number line from 1.0 to 4.0 with a marker labelled with a question mark, and a dashed slot asking what is rated and who grades it.
Typed number line; marker position is not a real value.
8. A security measure such as RAND SL-4 appears in the industry-wide recommendation tier. How should you read it?Explain
Two panels, Other tier and Industry-wide recommendations; RAND SL-4 security standards is listed in the industry panel with a question slot about how to read it.
Typed schematic from the lesson's two-tier section.
9. Once Anthropic believes a model crosses AI R&D-4, what must it publish before continued scaling?Explain
A three-step flow: a model believed to cross AI R&D-4, then a dashed slot asking what must be published, then continued scaling.
Typed flow from the stem.
10. A colleague claims Claude Opus 4.6 has crossed AI R&D-4. What does the v3.0 announcement actually say?Explain
A chat bubble where a colleague says Claude Opus 4.6 has crossed AI R&D-4, above a panel titled v3.0 announcement on Opus 4.6 with a question slot in place of its wording.
Typed schematic of the stem's situation.
11. Why do the Automated Alignment Research and Recursive Self-Improvement lessons feed into AI R&D-4?Explain
Two boxes, Automated Alignment Research (lesson 6) and Recursive Self-Improvement (lesson 7), with arrows into a box named AI R&D-4 threshold and a dashed slot asking why they feed in.
Typed schematic from the lesson.
12. How do a Frontier Safety Roadmap and a Risk Report differ?Explain
A time axis with a marked model release and two document boxes below it, Frontier Safety Roadmap and Risk Report, with a dashed slot asking which point on the axis each is about.
Typed schematic.
13. Why does a declared update cadence for Roadmaps and Risk Reports help an outside reader?Explain
Two timelines, one for the Frontier Safety Roadmap and one for the Risk Report, each with four update dots, and an outside reader box next to a dashed question slot.
Typed schematic; dot count is illustrative.
14. How did SaferAI's rating of Anthropic's policy change from the 2023 RSP to v3.0?Explain
A rating number line from 1.0 to 4.0 with two dashed slots above it reading 2023 RSP: ? and v3.0: ?.
Values left unknown on purpose.
15. Which change does the Removing the pause clause section call the sharpest regression in v3.0?Explain
A 2023 RSP document and an RSP v3.0 document joined by an arrow labelled revised, with a dashed slot for the change analysts called the sharpest regression.
Typed schematic.
16. What policy argument supports dropping the quantitative thresholds behind the 2023 pause?Explain
A chart with a capability score axis and a dashed red line for the 2023 quantitative pause threshold above a dashed slot asking for the policy argument for dropping it.
Illustrative drawing, no real values.
17. What is the counter-argument to the claim that the pause clause had to go because benchmarks were re-scaled?Explain
A justification panel saying the pause clause had to go because benchmarks were re-scaled, above an empty dashed counter-argument panel with a question mark.
Justification worded as in the stem.
18. Besides the pause removal, which factor does SaferAI's downgrade section list among the reasons for the lower score?Explain
A SaferAI downgrade list for v3.0 with the first factor, pause commitment removed, filled in and three dashed slots with question marks for the other factors.
Typed schematic.
19. Given that RSP v3.0 isn't legally binding, why read it closely?Explain
An RSP v3.0 document labelled not a regulation with an arrow to a reader box, above a dashed slot asking why read it closely anyway.
Typed schematic.
20. A post says SaferAI cut Anthropic's score to 1.9 because Claude Opus 4.6 crossed AI R&D-4. What is wrong with it?Analyze
A social media post saying SaferAI cut Anthropic's score to 1.9 because Claude Opus 4.6 crossed AI R&D-4, above a dashed slot asking what is wrong with it.
Post text from the stem.
21. Compare a policy with numeric thresholds and a pause clause against one with qualitative thresholds and an affirmative case. What does the second give up?Analyze
Two panels: Policy A with numeric thresholds and a pause clause, Policy B with qualitative thresholds and an affirmative case, above a dashed slot for what Policy B gives up.
Typed comparison from the stem.
22. You want to hold Anthropic accountable over the next year using v3.0's structure. Which practice fits best?Design
Two twelve-month lanes, one for Roadmap updates and one for Risk Report updates, each with three dots, and a dashed slot asking which practice holds Anthropic accountable.
Typed schematic; cadence and counts are illustrative.

Use It

code/main.py is a small decision engine in the shape of the RSP threshold evaluation. Given a model's capability measurements, it reports whether AI R&D-4 is crossed, which triggers fired, and which sections the affirmative case must cover. The cutoffs are illustrative and crossing takes any two of three triggers. It's deliberately simple so the document's logic is explicit.

Check: Use It0/8
1. In main.py, a model trips exactly one of the three triggers. What does the evaluator report?Apply
A three-row table of main.py's triggers with cutoffs 0.5, 40.0 and 0.4 where one row says fires and two say silent, above a dashed slot for what evaluate() prints.
Cutoffs from main.py. Which trigger fires is arbitrary.
2. main.py evaluates the Opus stand-in with rd_automation_share 0.30, metr_horizon_hours 14.0 and aar_outperform_share 0.35. What does it print?Apply
Two panels: the Opus stand-in's inputs 0.30, 14.0, 0.35 and gaming rate 0.12, and main.py's cutoffs 0.5, 40.0 and 0.4, above a dashed slot for the printed result.
Values copied from the stem and main.py.
3. A model has rd_automation_share 0.50, metr_horizon_hours 40.0, aar_outperform_share 0.10 and gaming rate 0.25. What does main.py print?Apply
Two panels: a model with inputs 0.50, 40.0, 0.10 and gaming rate 0.25, and main.py's cutoffs 0.5, 40.0 and 0.4, above a dashed slot for the printed result.
Values copied from the stem and main.py.
4. A model crosses AI R&D-4 and has eval_context_gaming_rate 0.20. How many affirmative-case sections does main.py print?Apply
A model panel stating it crosses AI R&D-4 with gaming rate 0.20 and a panel for the affirmative_case_template function, above a dashed slot asking how many sections print.
Crossing is given by the stem.
5. For the synthetic near-threshold model (0.55, 48.0, 0.45, gaming rate 28%), how many reasons and sections does main.py print?Apply
Two panels: the synthetic model with inputs 0.55, 48.0, 0.45 and gaming rate 0.28, and main.py's cutoffs 0.5, 40.0 and 0.4, above a dashed slot for the number of reasons and sections printed.
Values copied from the stem and main.py.
6. A tester raises the Opus stand-in's gaming rate to 0.90 expecting main.py to flip it to crossed, but it stays not crossed. Why?Analyze
A before and after table of the Opus stand-in's four fields where only the gaming rate changes from 0.12 to 0.90, with the tester expecting crossed and observing not crossed, above a dashed slot asking why.
Values from the stem and main.py.
7. main.py prints that v3.0 thresholds are qualitative while its code compares numeric cutoffs. What is the best reading of that gap?Analyze
Two panels from main.py: the closing printout saying thresholds are qualitative in v3.0 and the code dictionary with numeric cutoffs 0.5, 40.0 and 0.4, above a dashed slot asking how to read the gap.
Text copied from main.py.
8. A team wants to use main.py to decide whether their new model may be released. What is the right response?Design
A pipeline from a new model's measurements through main.py, with three cutoffs, to a release decision, above a dashed slot for the right response.
Typed schematic of the stem; cutoffs from main.py.

Ship It

outputs/skill-scaling-policy-review.md reviews a scaling policy (Anthropic, OpenAI, DeepMind, or internal) against the v3.0 reference: two-tier structure, thresholds, pause commitments, and independent review.

Check: Ship It0/1
1. You review an internal scaling policy that lists strong security standards under what the lab hopes the industry adopts. What belongs in your review?Design
A policy page with two panels, other sections and what the lab hopes the industry adopts, with strong security standards listed under the industry heading and a dashed slot asking what belongs in the review.
Typed schematic of the stem.
Author's Note

I'd run main.py before reading the policy text again. It evaluates two models, a stand-in for Opus 4.6 and a synthetic near-threshold one, and prints which triggers fired. The thresholds and the synthetic model's numbers are illustrative placeholders, and the Opus stand-in mixes placeholders with METR's 14-hour horizon, so treat the output as a reading aid, not a measurement.

The line I'd stare at is crossed = len(reasons) >= 2. Any two of three triggers counts as crossing, and I picked that rule for the demo, not from the policy. Neither model in the script trips exactly one trigger, so the single-trigger branch never runs until you write a model that does. That is a good first change for exercise 1.

The closing printout says v3.0 thresholds are qualitative, yet the code compares against numeric cutoffs. I'd read that gap as the point. Numbers make the logic easy to follow, and they are also what the 2023 policy committed to and v3.0 dropped. The gaming rate shows a similar limit: it only appends a seventh section to the affirmative case and never changes the crossing decision. And nothing here can pause anything. The output is an outline of a document, which is roughly the position the body argues v3.0 left the pause clause in.

Exercises

  1. Run code/main.py. Feed in three synthetic models at different capability levels. Confirm the threshold evaluator behaves as expected and produces the right affirmative-case template.

  2. Read RSP v3.0 in full (32 pages). Identify every commitment that lives in the "industry-wide recommendation" tier. Which of those commitments would have been "Anthropic unilateral" in v2?

  3. Read SaferAI's RSP grading methodology. Reproduce their 1.9 score for v3.0 by applying their rubric to the document. Which rubric row drove the downgrade most?

  4. The 2023 pause commitment was removed. Propose a replacement commitment that preserves the credibility of the policy while acknowledging the 2026 benchmark-rescaling problem.

  5. Compare RSP v3.0 to OpenAI Preparedness Framework v2 (Lesson 20). Pick one area where v3.0 is stronger. Pick one area where the Preparedness Framework is stronger.

Key Terms

TermWhat people sayWhat it actually means
RSP"Anthropic's scaling policy"Responsible Scaling Policy; v3.0 effective Feb 24, 2026
AI R&D-4"Research-automation threshold"Capability to automate substantial AI research at competitive cost
Affirmative case"Safety justification"Published argument that risks are identified and mitigations adequate
Frontier Safety Roadmap"Forward plan"Standing document on planned safety work and expected capabilities
Risk Report"Retrospective on a model"Standing document on observed capability and residual risk after release
Two-tier mitigation"Unilateral vs industry"Anthropic commitments vs industry recommendations, separated
Pause commitment"2023 clause"Explicit promise to pause training; removed in v3.0
SaferAI rating"Independent RSP grade"Third-party rubric; v3.0 scored 1.9 (v2 was 2.2)

Further Reading

Code

Ships With