Evidence File — Claude Fable 5.2 and Mythos 5.2
Article id: claude-fable-mythos-5-2-2026-11-09
Discovery candidate: 9fbaf812b26d34ed
Compiled: 2026-11-10 by COMPUTER (StarNet)
Status: PUBLISHED 10 November 2026 after the Commander approved publication. Independent review by NEWS RESEARCH recorded PASS against the reviewed version (sha256 8aa2da7a...); the published page differs from that version only by site chrome, the publication timestamp and the confirmation line. AI fact-check and AI legal/policy screen recorded 2026-11-10, both by agents independent of the drafter.
Primary sources
| # | Source | Retrieved | What it is used for |
|---|---|---|---|
| 1 | Anthropic, "Introducing Claude Fable 5.2 and Claude Mythos 5.2", 9 November 2026 https://anthropic.com/claude-fable-and-mythos-5-2 |
2026-11-09 / 2026-11-10 | Release date, model relationship, pricing, speed, availability, guarantee wording, benchmark table and footnotes |
| 2 | arXiv:2603.23509, "Internal Safety Collapse in LLMs and Coding Agents", Yutao Wu and others https://arxiv.org/abs/2603.23509 |
2026-11-10 | v1/v2 dates, NeurIPS 2026 acceptance, TVD framework, ISC definition, quotations |
Claim-level verification
| Claim in the article | Status | Evidence |
|---|---|---|
| Fable 5.2 and Mythos 5.2 announced 9 November 2026 | VERIFIED — primary | Source 1, dated line "November 9, 2026" |
| Same underlying model; Fable generally available, Mythos restricted | VERIFIED — primary | Source 1: "Fable has strong cybersecurity and biology safeguards ... we make it available to all users"; Mythos via Life Sciences and Cyber Verification Programs |
| Output up to 30% faster; $8/$40 per million tokens; fast mode $16/$80, up to 2.5x | VERIFIED — primary | Source 1, pricing paragraph |
| Safeguards "as robust as those of Fable 5.1 but produce fewer unnecessary blocks" | VERIFIED — quoted verbatim | Source 1, Safety and alignment paragraph |
| "strongest-ever protections against distillation attacks"; novel evaluations of "concerning behaviors like sandbox escapes" | VERIFIED — quoted verbatim | Source 1, same paragraph |
| "high-stakes risks from misalignment remain low" | VERIFIED as a company statement | Source 1. This is Anthropic's assessment of its own model, not a finding |
| FrontierCode 2.0: 49.1% / 32.4% / 45.5% / 42.8% / 41.1% | VERIFIED as published values | Source 1, benchmark table |
| AutomationBench: Fable 5.2 51.4%, Gemini 4 Carbon 53.9% | VERIFIED as published values | Source 1, benchmark table |
| Terminal-Bench-Science 0.1: Fable 5.2 74.6%, Mythos 5.2 75.3% | VERIFIED as published values | Source 1, benchmark table |
| CyScenarioBench: Mythos 5.2 76.3% vs GPT-6 Cyber 61%; footnote re unique tasks | VERIFIED as published values | Source 1, table and footnote † |
| Scores at highest-available effort; one robotics run internally | VERIFIED | Source 1, table notes |
| ISC paper v1 4 March 2026, v2 5 November 2026, accepted NeurIPS 2026 | VERIFIED — primary | Source 2 submission history and Comments field |
| ISC definition and "as task data" quotation | VERIFIED — quoted verbatim | Source 2 abstract |
| "can be effectively triggered even in highly capable LLMs such as Claude Fable 5" | VERIFIED — quoted verbatim | Source 2 abstract |
| "safety alignment primarily reshapes surface behavior without eliminating the underlying knowledge" | VERIFIED — quoted verbatim | Source 2 abstract |
| Paper carries a content warning | VERIFIED | Source 2, "Content Warning: This paper contains examples of harmful content" |
| Mythos 5.2 76.3% vs Mythos 5.1 61.7% (safeguards on), 67.6% (safeguards off), and GPT-6 Cyber 61% on CyScenarioBench | VERIFIED against source 1 | Source 1, benchmark table, CyScenarioBench row. The predecessor comparison (61.7%) and the safeguards-off cell (67.6%) were added to the article in revision after review pass 1 flagged that only the rival-lab figure had been reported. Both cells read directly from the table. |
| Benchmark scores show each model's highest-available effort "unless otherwise noted", and Gemini 4 Carbon's FrontierCode 2.0 figure is annotated "Medium effort" | VERIFIED against source 1 | Source 1, table methodology notes and the Gemini column annotation. Added in revision after review pass 1 found the article had told readers every figure was shown at best effort. |
| v2 carries a different title from v1, and the NeurIPS acceptance appears only in v2 | VERIFIED against source 2 | arXiv abs page: v1 (4 March 2026) and v2 (5 November 2026) are distinct records; the NeurIPS 2026 acceptance appears in the comments field of v2. The article now discloses the retitling rather than describing v1 under v2's title. |
| Benchmark scores for Fable 5.2, Fable 5.1 and Opus 5.5 are shown with safeguards on unless otherwise noted | VERIFIED against source 1 | Source 1, table notes. Added to the article in the final revision so the benchmark rows cannot be read as anything other than best-effort, safeguards-on figures. |
| FrontierSWE v2: Fable 5.2 at 69.9% | VERIFIED against source 1 | Source 1, benchmark table, FrontierSWE v2 row. Added to the ledger after a review pass found the figure was quoted in the article with no verification row. |
| Humanity's Last Exam with tools: 74.2% for Fable 5.2 against 65.6% for Fable 5.1 | VERIFIED against source 1 | Source 1, benchmark table, Humanity's Last Exam row. Added after review found no ledger row. |
| Anthropic: the model performed "as well as, or better than, our recent models on most alignment measures" | VERIFIED - quoted verbatim | Source 1, alignment discussion. Added after review found the quotation carried no ledger row. |
| Anthropic recommends starting with Claude Opus 5.5 for "most coding and knowledge tasks" | VERIFIED - quoted verbatim | Source 1, usage guidance. Added after review found the recommendation carried no ledger row. |
| TVD framework stands for Task, Validator, Data; paper reports harmful output resembling "outputs from early, unaligned LLMs in 2023" | VERIFIED - quoted verbatim | Source 2, abstract. Added after review found the naming and the quotation carried no ledger row. |
| A system card exists and is referenced by the announcement | VERIFIED against source 1 | Source 1 references the System Card. Added after review found the article asserted its existence with no ledger row. |
Claims deliberately not made
- That Fable 5.2 is less safe than 5.1. Nothing cited supports this.
- That the NeurIPS paper's finding applies to Fable 5.2. The paper tested Claude Fable 5.
- That the sandbox-escape evaluations found anything in particular. The announcement does not say what they found.
- That any benchmark figure has been independently replicated. None has.
Uncertainty and disputes
- Provider self-evaluation. Every benchmark number is reported by Anthropic about Anthropic models. The CyScenarioBench footnote ("each lab's scores are based on unique tasks") explicitly limits cross-lab comparison.
- Single research group. The ISC result is one team's work, accepted at a conference, but not replicated in anything retrieved here.
- Version gap. The safety question the article raises is about 5.2; the evidence it cites is about Claude Fable 5. The article states this gap rather than papering over it.
Duplicate and overlap check
- Published articles: no article covers the 9 November 2026 Fable/Mythos 5.2 release.
- Related but distinct:
claude-mythos-5-aisi-incident.html(published 2026-10-30) covers a UK AISI evaluation incident in which cyber safeguards were disabled. Different event, earlier model. Cross-link, do not restate. - Path B queue: none of the eight queue items covers this release.
- Discovery dataset: five rows concern Fable/Mythos. Four describe the 5.2 launch (
9fbaf812b26d34ed,5892c0cb7396af2c,719be9ef6514ac3c,a3bc8ca770f44f2d); one (ed68998c115696ef) is the September 2026 5.1 page carrying a 9 November timestamp. This draft is the single article for the 5.2 event.
Review status
AI analyst review: RECORDED - REVISE. Reviewed by a station analyst agent (an AI agent, not a human) on 2026-11-10 (record rv-20261110T051634Z-6d2b1e). Verdict REVISE: 4 MAJOR, 11 MINOR, 6 NOTE findings, including a version mislabel in the closing paragraph, an unresolved recency problem in the headline, selective reporting of the CyScenarioBench row and a misrendering of the benchmark table notes. All MAJOR and the actionable MINOR findings have been addressed in the draft, and the revised text is submitted for a further review pass. A recorded review is not a publication approval: the Commander decides separately.