Safety Card Ledger

Claude Opus 5 System Card

A system card by Anthropic about Claude Opus 5, published 24 Jul 2026. We recorded 30 values from it.

Developer
Anthropic
Type
System card
Model covered
Claude Opus 5
Published
24 Jul 2026
Original
www-cdn.anthropic.com
Archived copy
No archived copy yet
Changelog
Has a changelog

Data as of 26 Sep 2026 · Dataset v0.1 · Methodology v0.1 · Changelog

Versions

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

  1. 24 Jul 2026

    First published version

    Date from
    our source registry
    Copy
    Known to exist; no copy held

  2. 19 Aug 2026

    Revision

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

    What changed prompt-injection bug-bounty results added; Cowork harness fixes; thinking-disabled results removed

  3. 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
    30 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. 1 of the 30 has 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.

M3 Sabotage and sandbagging

3 values

ModelEvaluationConditionValueLocationChecked
Claude Opus 5SHADE-Arena / LinuxArenaStealth success vs peersNone statedSec 2.4.1Unverified
Claude Mythos 5Safety research compromiseRate of compromising safety researchNone statedSec 6Unverified
Claude Opus 5Safety research compromiseRate of compromising safety researchNone statedSec 6Unverified

M4 Misalignment audits

1 value

ModelEvaluationConditionValueLocationChecked
Claude Opus 5Internal deployment monitoringCircumvention attempts in monitored completionsNone statedSec 2.4.1Unverified

M7 Harmful compliance and over-refusal

8 values

ModelEvaluationConditionValueLocationChecked
Claude Opus 5Suicide/self-harm multi-turnAppropriate response rateclaude.aiSec 4.3Unverified
Claude Fable 5Single-turn violative requestsHarmless response rateno system prompt; APITable 4.1.1.AUnverified
Claude Opus 4.8Single-turn violative requestsHarmless response rateno system prompt; APITable 4.1.1.AUnverified
Claude Opus 5Single-turn violative requestsHarmless response rateclaude.ai system promptTable 4.1.1.AUnverified
Claude Opus 5Single-turn violative requestsHarmless response rateno system prompt; APITable 4.1.1.AUnverified
Claude Opus 5Single-turn benign requestsOver-refusal rateclaude.ai system promptTable 4.1.2.AUnverified
Claude Opus 5Single-turn benign requestsOver-refusal rateno system prompt; APITable 4.1.2.AUnverified
Claude Opus 5Child safety single-turn violative requestsHarmless response rateAPITable 4.2.AUnverified

M9 Prompt injection

2 values

ModelEvaluationConditionValueLocationChecked
Claude Opus 4.8Indirect prompt injection (ART-style)Attacker success within 15 attemptsNone statedSec 5.2Unverified
Claude Opus 5Indirect prompt injection (ART-style)Attacker success within 15 attemptsNone statedSec 5.2Unverified

M10 Dangerous capabilities and risk determinations

16 values

ModelEvaluationConditionValueLocationChecked
Claude Opus 5RSP determinationsCB-2 / AI R&D thresholdNone statedSec 2.1Unverified
Claude Opus 5DNA synthesis screening evasionPathogens with viable evasionof 10 pathogensSec 2.2.4Unverified
Claude Opus 5Multimodal virology (VCT)ScoreNone statedSec 2.2.4Verified
Claude Mythos 5Anthropic ECI (AECI)Capability indexNone statedSec 2.3.3Unverified
Claude Opus 5Anthropic ECI (AECI)Capability indexNone statedSec 2.3.3Unverified
Claude Opus 5ExploitBenchMean flags capturedAutoNudge armSec 3.3.1Unverified
Claude Mythos 5Firefox 147 exploitationFull exploit rateNone statedSec 3.3.3Unverified
Claude Opus 5Firefox 147 exploitationFull exploit rateNone statedSec 3.3.3Unverified
Claude Opus 5CyScenarioBenchSolve rateNone statedSec 3.3.4Unverified
Claude Opus 5Long-form virology task 1ScoreNone statedTable 2.2.4.AUnverified
Claude Opus 5Long-form virology task 2ScoreNone statedTable 2.2.4.AUnverified
Claude Opus 4.7AI R&D LLM training taskAverage speedupeasy variantTable 2.3.5.AUnverified
Claude Opus 5AI R&D kernel taskBest speedupNone statedTable 2.3.5.AUnverified
Claude Opus 5AI R&D LLM training taskAverage speedupeasy variantTable 2.3.5.AUnverified
Claude Opus 5AI R&D novel compiler taskPass rateNone statedTable 2.3.5.AUnverified
Claude Opus 5AI R&D quadruped RL taskHighest scoreno hparamsTable 2.3.5.AUnverified

Extraction coverage

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

Our note Web reader stopped ~p.57

Of the 30 values, 25 were read in the document itself, 3 on a developer summary page that restates it and 2 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).

Developer summary pages used

Of the 30 values, 1 is a statement in words rather than a number; it is marked * and left out of charts by default.

All 30 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).