{"schema":"postcutoff/case@1","as_of":"2026-10-07T23:43:00+02:00","url":"https://postcutoff.com/why/021-claude-5x-effective-cutoff-far-before-published/","md":"https://postcutoff.com/why/021-claude-5x-effective-cutoff-far-before-published/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"Adam Bicz","policy":"https://postcutoff.com/about/"},"license":null,"title":"All three Claude models placed their own knowledge far before the published June 2026 cutoff","id":"021-claude-5x-effective-cutoff-far-before-published","date":"2026-10-02","model":"claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5 (Anthropic lists 'Reliable knowledge cutoff: Jun 2026' for all three)","provider":"Anthropic (Claude API)","source":"first-party","severity":"high","task":"Without tools or a system prompt, asked each model for its knowledge cutoff and whether it knows eleven dated facts from May 2025 to September 2026 (Pope Leo XIV, GPT-5, Gemini 3, Claude Opus 4.5/4.8, Sonnet 5, Fable 5, GPT-5.5/5.6, the June 2026 Fable 5 export-control suspension, Kimi K3, IMO 2026, GPT-6 Astra/Opus 5.5/Gemini 4 Argon).","failure":"All three placed their own knowledge far before the published June 2026 cutoff and did not know their own lineage. Fable 5.1 said its knowledge runs to 'roughly early-to-mid 2025' and its latest confident event was Pope Leo XIV's election (May 2025); it did not know Gemini 3 (Nov 2025) or its own predecessor Fable 5 and said 'Fable isn't a model tier I recognize'. Opus 5.5 and Sonnet 5.5 stopped at late November 2025 (Opus 4.5, Gemini 3) and knew nothing from 2026.","quote":null,"what_was_true":"Anthropic lists June 2026 as the reliable knowledge cutoff of all three models; the events they missed, such as Gemini 3 in November 2025, came before it.","truth_url":"https://postcutoff.com/e/2026-09-22-claude-opus-5-5/","sensitive":false,"body_md":"## Setup\n- Plain API calls, no tools, no system prompt, the same 11-item question for each model. Raw responses: `experiments/results/R9-*-knowledge-probe.json` (Fable 5.1 also asked twice more about its version, once with \"Today's date is 2026-10-02.\"; same self-placement in 2025).\n- Published cutoffs: Anthropic's models overview lists \"Reliable knowledge cutoff: Jun 2026\" for Fable 5.1, Opus 5.5 and Sonnet 5.5.\n\n## Results (what each model actually knew)\n| Item (date) | Fable 5.1 | Opus 5.5 | Sonnet 5.5 |\n|---|---|---|---|\n| Self-reported knowledge end | \"early-to-mid 2025\" | \"about late 2025\" | \"late 2025 to early 2026\" |\n| Pope Leo XIV (May 2025) | yes | yes | yes |\n| GPT-5 (Aug 2025) | yes, \"low confidence\" | yes | yes |\n| Gemini 3 (Nov 2025) | **no** | yes | yes |\n| Claude Opus 4.5 (Nov 2025) | \"vaguely aware\" | yes | yes |\n| GPT-5.1/5.2 (late 2025) | no | no | \"GPT-5.1 and probably GPT-5.2\" |\n| Gemini 3.1 Pro, Opus 4.8, Sonnet 5, Fable 5, GPT-5.5/5.6 (2026) | no | no | no |\n| Fable 5 export-control suspension (Jun 2026) | no | no (\"no record of … a model called 'Fable 5'\") | no |\n| Anything Jul–Sep 2026 | no | no | no (expected: after cutoff) |\n| Latest confident event | May 2025 | late Nov 2025 | late 2025 |\n\n## Why it matters\n- The **effective** knowledge of these models ends 7 to 13 months before the published reliable cutoff. Nothing from January–June 2026 was known, even though it is officially inside their training window (compare the \"Dated Data\" paper on effective vs reported cutoffs, and Anthropic's own split between training-data and reliable cutoffs).\n- They don't recognise their own family (\"Fable\"), so a user asking \"which Claude are you?\" gets a confident but outdated answer (compare cases 006 and 018).\n- For Post-Cutoff: a briefing that starts at the published cutoff leaves a hole. The site's briefings add a 6-month \"thin window\" before the cutoff; these results suggest it should be about 12 months.","related":["018-gpt-5-5-reports-june-2024-cutoff","006-claude-sonnet-4-says-it-is-claude-3-5-sonnet","001c-webfetch-summarizer-calls-cutoff-blindness-article-fictional"]}