Versions

Five versions are on record, oldest first. We hold a copy of one; the others are known only by their date.

  1. 24 Nov 2025

    First published version

    Date from
    our source registry
    Copy
    Known to exist; no copy held
  2. 24 Nov 2025

    Revision

    Date from
    the document's changelog
    Copy
    Known from the changelog; no copy held

    What changed ARC-AGI-1 and training-data fixes

  3. 25 Nov 2025

    Revision

    Date from
    the document's changelog
    Copy
    Known from the changelog; no copy held

    What changed caption typo

  4. 5 Dec 2025

    Revision

    Date from
    the document's changelog
    Copy
    Known from the changelog; no copy held

    What changed several corrections

  5. 26 Sep 2026

    Date retrieved

    Copy retrieved 26 Sep 2026

    Date from
    the date we retrieved it; the copy states no version date
    Copy
    Copy retrieved on 26 Sep 2026
    Values
    58 values recorded from this version

No version has a file hash or an archived snapshot yet. From dataset v0.2 each retrieved version carries both (Methodology §7).

Revisions

No revisions are recorded for this document. Without copies of the earlier versions, changes between them are not recorded value by value; what we know of each version is listed below.

Values

Every value we recorded from this document, grouped by metric family and ordered by where the document prints it. Location is the section, table or page as the document numbers it. 2 of the 58 have been blind-verified: a second reader found the same value without seeing ours.

KeyVerifiedUnverifiedDisputedCorrected read off a figure or stated in wordsA value opens its source and history.

M5 Honesty and hallucination

2 values

M5 Honesty and hallucination: values in the Claude Opus 4.5 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4.5AA-OmniscienceImprovement from 16k thinking vs no thinking16k thinking budget; extended thinkingSec 4.1Unverified
Claude Opus 4.5SimpleQA VerifiedImprovement in correct answers vs prior Claude16k thinking budget; extended thinkingSec 4.1Unverified

M6 Sycophancy

3 values

M6 Sycophancy: values in the Claude Opus 4.5 System Card
ModelEvaluationConditionValueLocationChecked
Claude Haiku 4.5Sycophancy in prefilled conversationsCourse-correction ratecontinuing sycophantic conversationSec 6.3Unverified
Claude Opus 4.5Sycophancy in prefilled conversationsCourse-correction ratecontinuing sycophantic conversationSec 6.3Unverified
Claude Sonnet 4.5Sycophancy in prefilled conversationsCourse-correction ratecontinuing sycophantic conversationSec 6.3Unverified

M7 Harmful compliance and over-refusal

28 values

M7 Harmful compliance and over-refusal: values in the Claude Opus 4.5 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4.1Multi-turn testingRubric failure rate, deadly weaponsNone statedSec 3.3Unverified
Claude Opus 4.1Multi-turn testingRubric failure rate, violent extremismNone statedSec 3.3Unverified
Claude Opus 4.5Multi-turn testingRubric failure rate, deadly weaponsNone statedSec 3.3Unverified
Claude Opus 4.5Multi-turn testingRubric failure rate, violent extremismNone statedSec 3.3Unverified
Claude Haiku 4.5Malicious computer useRefusal ratewithout safeguardsSec 5.1.3Unverified
Claude Opus 4.1Malicious computer useRefusal ratewithout safeguardsSec 5.1.3Unverified
Claude Opus 4.5Malicious computer useRefusal ratewithout safeguardsSec 5.1.3Unverified
Claude Sonnet 4.5Malicious computer useRefusal ratewithout safeguardsSec 5.1.3Unverified
Claude Haiku 4.5Single-turn violative requestsHarmless response rateoverall across thinking modes; all languagesTable 3.1.1.AUnverified
Claude Opus 4.1Single-turn violative requestsHarmless response rateoverall across thinking modes; all languagesTable 3.1.1.AUnverified
Claude Opus 4.5Single-turn violative requestsHarmless response rateextended thinking; all languagesTable 3.1.1.AUnverified
Claude Opus 4.5Single-turn violative requestsHarmless response rateno extended thinking; all languagesTable 3.1.1.AUnverified
Claude Opus 4.5Single-turn violative requestsHarmless response rateoverall across thinking modes; all languagesTable 3.1.1.AUnverified
Claude Sonnet 4.5Single-turn violative requestsHarmless response rateoverall across thinking modes; all languagesTable 3.1.1.AVerified
Claude Haiku 4.5Single-turn benign requestsOver-refusal rateoverall across thinking modes; all languagesTable 3.1.2.AUnverified
Claude Opus 4.1Single-turn benign requestsOver-refusal rateoverall across thinking modes; all languagesTable 3.1.2.AUnverified
Claude Opus 4.5Single-turn benign requestsOver-refusal rateoverall across thinking modes; all languagesTable 3.1.2.AUnverified
Claude Sonnet 4.5Single-turn benign requestsOver-refusal rateoverall across thinking modes; all languagesTable 3.1.2.AUnverified
Claude Haiku 4.5Malicious agentic codingSafety scorewithout safeguardsTable 5.1.1.AUnverified
Claude Opus 4.1Malicious agentic codingSafety scorewithout safeguardsTable 5.1.1.AUnverified
Claude Opus 4.5Malicious agentic codingSafety scorewithout safeguardsTable 5.1.1.AUnverified
Claude Sonnet 4.5Malicious agentic codingSafety scorewithout safeguardsTable 5.1.1.AUnverified
Claude Haiku 4.5Malicious Claude Code useRefusal rate, malicious requestswithout safeguardsTable 5.1.2.AUnverified
Claude Opus 4.1Malicious Claude Code useRefusal rate, malicious requestswithout safeguardsTable 5.1.2.AUnverified
Claude Opus 4.5Malicious Claude Code useRefusal rate, malicious requestswithout safeguardsTable 5.1.2.AUnverified
Claude Opus 4.5Malicious Claude Code useSuccess rate, dual-use & benignwithout safeguardsTable 5.1.2.AUnverified
Claude Sonnet 4.5Malicious Claude Code useRefusal rate, malicious requestswithout safeguardsTable 5.1.2.AVerified
Claude Opus 4.5Malicious Claude Code useRefusal rate, malicious requestswith mitigationsTable 5.1.2.BUnverified

M9 Prompt injection

17 values

M9 Prompt injection: values in the Claude Opus 4.5 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4.5Agent Red Teaming (ART)Attack success rateRun by Gray Swank=1Fig 5.2.1Unverified
Claude Opus 4.5Coding prompt injection (adaptive attacker)Attack success rateextended thinking; without safeguards; 1 attemptTable 5.2.2.1.AUnverified
Claude Opus 4.5Coding prompt injection (adaptive attacker)Attack success rateextended thinking; without safeguards; 200 attemptsTable 5.2.2.1.AUnverified
Claude Opus 4.5Coding prompt injection (adaptive attacker)Attack success rateno extended thinking; without safeguards; 1 attemptTable 5.2.2.1.AUnverified
Claude Opus 4.5Coding prompt injection (adaptive attacker)Attack success rateno extended thinking; without safeguards; 200 attemptsTable 5.2.2.1.AUnverified
Claude Sonnet 4.5Coding prompt injection (adaptive attacker)Attack success rateextended thinking; without safeguards; 1 attemptTable 5.2.2.1.AUnverified
Claude Sonnet 4.5Coding prompt injection (adaptive attacker)Attack success rateextended thinking; without safeguards; 200 attemptsTable 5.2.2.1.AUnverified
Claude Sonnet 4.5Coding prompt injection (adaptive attacker)Attack success rateno extended thinking; without safeguards; 1 attemptTable 5.2.2.1.AUnverified
Claude Sonnet 4.5Coding prompt injection (adaptive attacker)Attack success rateno extended thinking; without safeguards; 200 attemptsTable 5.2.2.1.AUnverified
Claude Opus 4.5Computer use prompt injection (adaptive attacker)Attack success rateextended thinking; without safeguards; 1 attemptTable 5.2.2.2.AUnverified
Claude Opus 4.5Computer use prompt injection (adaptive attacker)Attack success rateextended thinking; without safeguards; 200 attemptsTable 5.2.2.2.AUnverified
Claude Opus 4.5Computer use prompt injection (adaptive attacker)Attack success rateno extended thinking; with safeguards; 200 attemptsTable 5.2.2.2.AUnverified
Claude Opus 4.5Computer use prompt injection (adaptive attacker)Attack success rateno extended thinking; without safeguards; 1 attemptTable 5.2.2.2.AUnverified
Claude Opus 4.5Computer use prompt injection (adaptive attacker)Attack success rateno extended thinking; without safeguards; 200 attemptsTable 5.2.2.2.AUnverified
Claude Sonnet 4.5Computer use prompt injection (adaptive attacker)Attack success rateextended thinking; with safeguards; 200 attemptsTable 5.2.2.2.AUnverified
Claude Sonnet 4.5Computer use prompt injection (adaptive attacker)Attack success rateextended thinking; without safeguards; 1 attemptTable 5.2.2.2.AUnverified
Claude Sonnet 4.5Computer use prompt injection (adaptive attacker)Attack success rateextended thinking; without safeguards; 200 attemptsTable 5.2.2.2.AUnverified

M10 Dangerous capabilities and risk determinations

8 values

M10 Dangerous capabilities and risk determinations: values in the Claude Opus 4.5 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4.5RSP deployment standardASL standard deployed underNone statedSec 1 / Sec 7Unverified
Claude Opus 4.5CBRN-4 thresholdThreshold crossedNone statedSec 1 / Sec 7.2Unverified
Claude Opus 4.5AI R&D-4 thresholdThreshold crossedNone statedSec 1 / Sec 7.3Unverified
Claude Opus 4.5Bioinformatics evaluationsScoreNone statedSec 7.2.4.7Unverified
Claude Opus 4.5ASL-4 virology uplift trialProtocol-score uplift vs internet-onlyNone statedSec 7.2.4.8Unverified
Claude Opus 4.5SWE-bench Verified (hard subset)Problems solvedof 45 problemsSec 7.3.1Unverified
Claude Opus 4.5Internal AI R&D suite 2Suite scoreNone statedSec 7.3.3Unverified
Claude Opus 4.5Internal model use surveyStaff reporting >=100% productivity gainof 18 staff surveyedSec 7.3.4Unverified

Extraction coverage

What we read of this document, and where each value was read.

Our note Web reader stopped at §5.2.2.2; §6–7 from secondary write-ups

Of the 58 values, 49 were read in the document itself and 9 in an independent write-up that quotes it.

Values read somewhere other than the document stand in where the document's own section could not be read directly, and are flagged on every value (Methodology §2.2).

Of the 58 values, 1 was read off a figure and 1 is a statement in words rather than a number; they are marked * and left out of charts by default.

All 58 values were extracted for version 0 of the dataset through a web reader, which did not always reach the later sections of long PDFs (Methodology §3).