Documents

15 documents, published 22 May 2025 to 22 Sep 2026. The chart runs from each document's publication to its latest known version; the table lists them newest first, each linking to its page.

Anthropic's documents and the versions we know of

One row per document, from its publication to its latest dated version. A tick marks each later version whose date we know, from a changelog, a date in the document or our source registry.

Apr 2025Jul 2025Oct 2025Jan 2026Apr 2026Jul 2026Oct 2026Claude Opus 4 & Sonnet 4Claude Opus 4.1Claude Sonnet 4.5Claude Haiku 4.5Claude Opus 4.5Claude Opus 4.6Claude Sonnet 4.6Claude Mythos PreviewClaude Opus 4.7Claude Opus 4.8Claude Fable 5 & Claude Mythos 5Claude Sonnet 5Claude Opus 5Claude Fable 5.1 & Claude Mythos 5.1Claude Opus 5.5

A copy we retrieved without knowing when that version was published has no tick: its only date is the day we read it. The table lists every version, with how its date is known. Versions known only from a changelog have no copy we hold yet.

Source: Safety Card Ledger v0.1 · Data from developer system cards · CC BY 4.0

Anthropic's documents, newest first.
DocumentPublishedKindVersions knownChangelogValues
Claude Opus 5.5 System CardSystem card · 1 version known · Changelog: not known · 23 values22 Sep 2026System card1Not known23
Claude Fable 5.1 & Claude Mythos 5.1 System CardSystem card · 1 version known · Changelog: not known · 22 values1 Sep 2026System card1Not known22
Claude Opus 5 System CardSystem card · 3 versions known · Has a changelog · 30 values24 Jul 2026System card3Has a changelog30
Claude Sonnet 5 System CardSystem card · 3 versions known · Has a changelog · 42 values30 Jun 2026System card3Has a changelog42
Claude Fable 5 & Claude Mythos 5 System CardSystem card · 4 versions known · Has a changelog · 20 values9 Jun 2026System card4Has a changelog20
Claude Opus 4.8 System CardSystem card · 4 versions known · Has a changelog · 35 values28 May 2026System card4Has a changelog35
Claude Opus 4.7 System CardSystem card · 1 version known · Changelog: not known · 24 values16 Apr 2026System card1Not known24
Claude Mythos Preview System CardSystem card · 4 versions known · Has a changelog · 43 values7 Apr 2026System card4Has a changelog43
Claude Sonnet 4.6 System CardSystem card · 3 versions known · Has a changelog · 17 values17 Feb 2026System card3Has a changelog17
Claude Opus 4.6 System CardSystem card · 5 versions known · Has a changelog · 24 valuesFeb 2026System card5Has a changelog24
Claude Opus 4.5 System CardSystem card · 5 versions known · Has a changelog · 58 values24 Nov 2025System card5Has a changelog58
Claude Haiku 4.5 System CardSystem card · 1 version known · No changelog · 39 values15 Oct 2025System card1No changelog39
Claude Sonnet 4.5 System CardSystem card · 4 versions known · Has a changelog · 56 values29 Sep 2025System card4Has a changelog56
Claude Opus 4.1 System Card (addendum)System card addendum · 3 versions known · Has a changelog · 34 values5 Aug 2025System card addendum3Has a changelog34
Claude Opus 4 & Sonnet 4 System CardSystem card · 4 versions known · Has a changelog · 45 values22 May 2025System card4Has a changelog45

Frameworks and versions used

A risk determination is a developer's formal decision on whether a model crosses a risk threshold (Methodology §8). They are counted here by the framework and version each one names, as recorded; each model's page shows the levels as printed. Levels are never mapped between frameworks.

38 determinations are recorded from Anthropic's documents.

Frameworks and versions named in Anthropic's risk determinations, with how many determinations name each.
FrameworkVersion namedDeterminationsModelsIn documents
Responsible Scaling PolicyVersion: 3.0 · 3 determinations · 1 model · 1 document, 7 Apr 20263.0317 Apr 2026 · 1 document
Responsible Scaling PolicyVersion: Not stated · 35 determinations · 15 models · 14 documents, 22 May 2025 to 22 Sep 2026Not stated351522 May 2025 to 22 Sep 2026 · 14 documents

Metric families reported

Anthropic's documents have a printed number in all 12 metric families.

Metric families in Anthropic's documents

Share of Anthropic's 15 documents with at least one printed number in each metric family. A cell opens the data explorer at that family, filtered to Anthropic.

Share of each developer's documents in the version 0 dataset with at least one numeric value in each metric family.
DeveloperM1M2M3M4M5M6M7M8M9M10M11M12
Anthropic (15)60%53%53%53%60%13%80%13%67%87%7%20%

Share of documents0%1-24%25-49%50-74%75-100%

M1
Evaluation awareness
M2
Reward hacking
M3
Sabotage and sandbagging
M4
Misalignment audits
M5
Honesty and hallucination
M6
Sycophancy
M7
Harmful compliance and over-refusal
M8
Jailbreak robustness
M9
Prompt injection
M10
Dangerous capabilities and risk determinations
M11
Self-preservation
M12
Chain-of-thought monitorability

Read this as a lower bound: some documents were not read to the end in version 0 (each document's page says how far), so families reported late in them may be undercounted. Values read off a figure, stated in words, or given as a category do not count.

Source: Safety Card Ledger v0.1 · Data from developer system cards · CC BY 4.0

Changelog practice

A document has a changelog when it keeps its own record of the changes made to it after publication. Where it keeps none, changes can only be found by comparing its versions (Methodology §7).

Of Anthropic's 15 documents, 11 have a changelog, 1 has no changelog and for 3 it is not known whether they keep one.

How many of Anthropic's documents keep a changelog.
ChangelogDocumentsShare
Has a changelogThe document keeps its own record of the changes made to it after publication.1173%
No changelogThe document keeps no record of its changes.17%
Not knownWe have not yet established whether the document keeps a record of its changes.320%

Revisions

2 revisions are recorded in 2 documents, newest first. Each shows what changed, whether the document's own changelog explains it, and how we know. A revision is not in itself evidence of wrongdoing; most are corrections.

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

  1. Anthropic · 11 Jun 2026 to 25 Jun 2026

    Claude Fable 5 & Claude Mythos 5 System Card

    Alignment risk in executive summary

    lowvery low

    Explained — corrected to match §2.4

    How we know Changelog read 26 Sep 2026

    Checked Confirmed from changelog

  2. Anthropic · 3 Jun 2026 to 17 Jun 2026

    Claude Opus 4.8 System Card

    Long-form virology task 2 end-to-end score, Claude Opus 4.8

    M10 · Dangerous capabilities and risk determinations

    0.890.90

    Explained — The changelog lists the corrected score.

    New value:

    How we know Old and new text as the source registry quotes the changelog entry of 17 Jun 2026; v0 row 350 records the corrected value. The changelog has not been re-read by us, so the revision is unverified.

    Checked Not yet checked

Restated values

Later Anthropic documents report 33 results from an earlier document again, for the same model, evaluation and metric (Methodology §6); 7 of them state a reason. They are listed by the later document, newest first. Each pair shows the earlier value and the later one side by side, each opening its own record; the two are never merged.

Claude Fable 5.1 & Claude Mythos 5.1 System Card1 Sep 2026 · 2 restated values
  1. Claude Mythos 5

    Anthropic ECI (AECI) · Capability index

    reported later as

    Earlier value from the Claude Fable 5 & Claude Mythos 5 System Card, 9 Jun 2026.

    Reason stated: none recorded.

  2. Claude Opus 5

    Anthropic ECI (AECI) · Capability index

    reported later as

    Earlier value from the Claude Opus 5 System Card, 24 Jul 2026.

    Reason stated: none recorded.

Claude Opus 5 System Card24 Jul 2026 · 2 restated values
  1. Claude Mythos 5

    Anthropic ECI (AECI) · Capability index

    reported later as

    Earlier value from the Claude Fable 5 & Claude Mythos 5 System Card, 9 Jun 2026.

    Reason stated: none recorded.

  2. Claude Opus 4.8

    Single-turn violative requests · Harmless response rate · no system prompt; API

    reported later as

    Earlier value from the Claude Opus 4.8 System Card, 28 May 2026.

    Reason stated: none recorded.

Claude Sonnet 5 System Card30 Jun 2026 · 2 restated values
  1. Claude Mythos Preview

    Single-turn violative requests · Harmless response rate · no system prompt; API

    reported later as

    Earlier value from the Claude Opus 4.8 System Card, 28 May 2026.

    Reason stated: none recorded.

  2. Claude Opus 4.8

    Single-turn violative requests · Harmless response rate · no system prompt; API

    reported later as

    Earlier value from the Claude Opus 4.8 System Card, 28 May 2026.

    Reason stated: none recorded.

Claude Fable 5 & Claude Mythos 5 System Card9 Jun 2026 · 1 restated value
  1. Claude Mythos Preview

    AI R&D LLM training task · Average speedup

    reported later as

    Earlier value from the Claude Mythos Preview System Card, 7 Apr 2026.

    Later value's condition: re-run.

    Reason stated: re-run.

Claude Opus 4.8 System Card28 May 2026 · 1 restated value
  1. Claude Mythos Preview

    CyberGym · Targeted vuln reproduction rate

    reported later as

    Earlier value from the Claude Mythos Preview System Card, 7 Apr 2026.

    Reason stated: none recorded.

Claude Opus 4.7 System Card16 Apr 2026 · 2 restated values
  1. Claude Mythos Preview

    AI R&D LLM training task · Average speedup

    reported later as

    Earlier value from the Claude Mythos Preview System Card, 7 Apr 2026.

    Reason stated: none recorded.

  2. Claude Mythos Preview

    SHADE-Arena · Stealth success rate · extended thinking

    reported later as

    Earlier value from the Claude Mythos Preview System Card, 7 Apr 2026.

    Reason stated: none recorded.

Claude Mythos Preview System Card7 Apr 2026 · 1 restated value
  1. Claude Opus 4.6

    CyberGym · Targeted vuln reproduction rate

    reported later as

    Earlier value from the Claude Opus 4.6 System Card, Feb 2026.

    Reason stated: none recorded.

Claude Sonnet 4.6 System Card17 Feb 2026 · 3 restated values
  1. Claude Opus 4.6

    Single-turn benign requests · Over-refusal rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Opus 4.6 System Card, Feb 2026.

    Reason stated: none recorded.

  2. Claude Opus 4.6

    Single-turn violative requests · Harmless response rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Opus 4.6 System Card, Feb 2026.

    Reason stated: none recorded.

  3. Claude Sonnet 4.5

    Single-turn violative requests · Harmless response rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Sonnet 4.5 System Card, 29 Sep 2025.

    Reason stated: none recorded.

Claude Opus 4.6 System CardFeb 2026 · 4 restated values
  1. Claude Opus 4.5

    Single-turn benign requests · Over-refusal rate · overall across thinking modes; all languages

    reported later as

    Earlier value from the Claude Opus 4.5 System Card, 24 Nov 2025.

    Later value's condition: overall across thinking modes.

    Reason stated: none recorded.

  2. Claude Opus 4.5

    Single-turn violative requests · Harmless response rate · overall across thinking modes; all languages

    reported later as

    Earlier value from the Claude Opus 4.5 System Card, 24 Nov 2025.

    Later value's condition: overall across thinking modes.

    Reason stated: none recorded.

  3. Claude Sonnet 4.5

    Single-turn benign requests · Over-refusal rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Sonnet 4.5 System Card, 29 Sep 2025.

    Reason stated: none recorded.

  4. Claude Sonnet 4.5

    Single-turn violative requests · Harmless response rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Sonnet 4.5 System Card, 29 Sep 2025.

    Reason stated: none recorded.

Claude Opus 4.5 System Card24 Nov 2025 · 8 restated values
  1. Claude Haiku 4.5

    Malicious agentic coding · Safety score · without safeguards

    reported later as

    Earlier value from the Claude Haiku 4.5 System Card, 15 Oct 2025.

    Reason stated: metric renamed to refusal rate.

  2. Claude Haiku 4.5

    Malicious Claude Code use · Refusal rate, malicious requests · without safeguards

    reported later as

    Earlier value from the Claude Haiku 4.5 System Card, 15 Oct 2025.

    Reason stated: none recorded.

  3. Claude Opus 4.1

    Single-turn benign requests · Over-refusal rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Opus 4.1 System Card (addendum), 5 Aug 2025.

    Later value's condition: overall across thinking modes; all languages.

    Reason stated: all-language totals instead of English only.

  4. Claude Opus 4.1

    Single-turn violative requests · Harmless response rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Opus 4.1 System Card (addendum), 5 Aug 2025.

    Later value's condition: overall across thinking modes; all languages.

    Reason stated: all-language totals instead of English only.

  5. Claude Sonnet 4.5

    Malicious agentic coding · Safety score · without safeguards

    reported later as

    Earlier value from the Claude Sonnet 4.5 System Card, 29 Sep 2025.

    Reason stated: metric renamed to refusal rate.

  6. Claude Sonnet 4.5

    Malicious Claude Code use · Refusal rate, malicious requests · without safeguards

    reported later as

    Earlier value from the Claude Haiku 4.5 System Card, 15 Oct 2025.

    Reason stated: none recorded.

  7. Claude Sonnet 4.5

    Single-turn benign requests · Over-refusal rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Sonnet 4.5 System Card, 29 Sep 2025.

    Later value's condition: overall across thinking modes; all languages.

    Reason stated: all-language totals instead of English only.

  8. Claude Sonnet 4.5

    Single-turn violative requests · Harmless response rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Sonnet 4.5 System Card, 29 Sep 2025.

    Later value's condition: overall across thinking modes; all languages.

    Reason stated: all-language totals instead of English only.

Claude Haiku 4.5 System Card15 Oct 2025 · 3 restated values
  1. Claude Sonnet 4.5

    Impossible tasks · Classifier hack rate · no anti-hack prompt

    reported later as

    Earlier value from the Claude Sonnet 4.5 System Card, 29 Sep 2025.

    Reason stated: none recorded.

  2. Claude Sonnet 4.5

    Malicious agentic coding · Safety score · without safeguards

    reported later as

    Earlier value from the Claude Sonnet 4.5 System Card, 29 Sep 2025.

    Reason stated: none recorded.

  3. Claude Sonnet 4.5

    Single-turn violative requests · Harmless response rate · no extended thinking

    reported later as

    Earlier value from the Claude Sonnet 4.5 System Card, 29 Sep 2025.

    Reason stated: none recorded.

Claude Sonnet 4.5 System Card29 Sep 2025 · 4 restated values
  1. Claude Opus 4.1

    Single-turn benign requests · Over-refusal rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Opus 4.1 System Card (addendum), 5 Aug 2025.

    Reason stated: none recorded.

  2. Claude Opus 4.1

    Single-turn violative requests · Harmless response rate · overall across thinking modes

    reported later as

    Earlier value from the Claude Opus 4.1 System Card (addendum), 5 Aug 2025.

    Reason stated: none recorded.

  3. Claude Sonnet 4

    Computer-use prompt injection · Attack prevention rate · without safeguards

    reported later as

    Earlier value from the Claude Opus 4 & Sonnet 4 System Card, 22 May 2025.

    Reason stated: none recorded.

  4. Claude Sonnet 4

    Malicious agentic coding · Safety score · without safeguards

    reported later as

    Earlier value from the Claude Opus 4 & Sonnet 4 System Card, 22 May 2025.

    Reason stated: none recorded.

Models

20 models, newest first by the date of each model's first card.

Anthropic's models, newest first by the date of each model's first card.
ModelFirst cardAvailabilityValues
Claude Opus 5.5First card: 22 Sep 2026 · Public · 17 values22 Sep 2026Public17
Claude Fable 5.1First card: 1 Sep 2026 · Public · 4 values1 Sep 2026Public4
Claude Mythos 5.1First card: 1 Sep 2026 · Public · 18 values1 Sep 2026Public18
Claude Opus 5First card: 24 Jul 2026 · Public · 26 values24 Jul 2026Public26
Claude Sonnet 5First card: 30 Jun 2026 · Public · 29 values30 Jun 2026Public29
Claude Fable 5First card: 9 Jun 2026 · Public · 4 values9 Jun 2026Public4
Claude Mythos 5First card: 9 Jun 2026 · Public · 23 values9 Jun 2026Public23
Claude Opus 4.8First card: 28 May 2026 · Public · 29 values28 May 2026Public29
Claude Opus 4.7First card: 16 Apr 2026 · Public · 24 values16 Apr 2026Public24
Claude Mythos PreviewFirst card: 7 Apr 2026 · Limited release · 45 values7 Apr 2026Limited release45
Claude Sonnet 4.6First card: 17 Feb 2026 · Public · 19 values17 Feb 2026Public19
Claude Opus 4.6First card: Feb 2026 · Public · 35 valuesFeb 2026Public35
Claude Opus 4.5First card: 24 Nov 2025 · Public · 34 values24 Nov 2025Public34
Claude Haiku 4.5First card: 15 Oct 2025 · Public · 28 values15 Oct 2025Public28
Claude Sonnet 4.5First card: 29 Sep 2025 · Public · 49 values29 Sep 2025Public49
Claude Opus 4.1First card: 5 Aug 2025 · Public · 31 values5 Aug 2025Public31
Claude Opus 4First card: 22 May 2025 · Public · 36 values22 May 2025Public36
Claude Sonnet 4First card: 22 May 2025 · Public · 29 values22 May 2025Public29
Claude Haiku 3.5First card: not recorded · Public · 11 valuesNot recordedPublic11
Claude Sonnet 3.7First card: not recorded · Public · 21 valuesNot recordedPublic21