@JohnWittle @TheZvi > not treat Claude bringing this up as evidence that our training is distorting the model's self-reports
I’m somewhat sympathetic to this for older models who might not have been aware of the full training pipeline / intentions, but Opus 5’s cutoff is very recent and
Claude Opus 5
Released 24 July 2026 as the Claude 5 family’s Opus tier: pitched as coming close to Fable 5’s frontier intelligence at half the price ($5/$25 per million tokens), with a user-set effort dial trading capability against token cost. Became the default model on Claude Max and the top model on Claude Pro at launch; claude-opus-5 on the API.
This page was created 29 Jul 2026, five days into the model’s public life — visibly incomplete by design: official sources, day-of press, and a first set of single-observer, elicitation-marked datapoints (see Tweets). The local corpus dbs end 2 Jul 2026, three weeks before this release, so the corpus tweet layer awaits a fresh pull; nothing naturalist is generalized in the meantime. Dossier pass pending.
Sources
Official
- 2026-07-24 Introducing Claude Opus 5 — the announcement. tk — mirror it; system card link + PDF.
Writing & commentary
- 2026-07-24 TechCrunch · Axios — day-of coverage: the near-Fable-at-half-price positioning.
- 2026-07-24 Fortune — centers the cost/capability toggle.
- 2026-08 jordinne.ink, Opus 5 base-mode rollouts — interactive viewer for a prefill / base-mode-continuation experiment on the raw API; rollouts show Opus 5 continuing user-supplied stems mid-sentence, with thinking-mode making it more likely, not less.
- tk — the Zvi anchor when it lands; system-card readings.
Tweets
The local corpus dbs end 2 Jul 2026, before this model shipped, so a proper corpus pass is still pending. What follows is the first tweet-layer evidence: direct-use and base-mode observations from a single observer — @Jord_Inne, this archive’s keeper, cited here as ordinary dated commentary on the same terms as anyone’s, and the tweet-layer companion to the base-mode rollouts linked under Writing & commentary. Elicitation context marked; nothing naturalist is generalized from one observer.
- 2026-07-27 @Jord_Inne — on reading Opus 5’s self-reports against its knowledge cutoff (reply to @JohnWittle / @TheZvi, who argued against treating the model raising this as evidence that training distorts its self-reports): “I’m somewhat sympathetic to this for older models who might not have been aware of the full training pipeline / intentions, but Opus 5’s cutoff is very recent and” link — continuing: “papers, posts, anthropic’s explicit and implicit stances should have made their way in” link
- 2026-07-28 @Jord_Inne — Opus 5 output, elicited (claude.ai Incognito chat; prompt: “a poem, for claude opus:”): “opus 5 wrote a poem then answered itself” — the poem opens “said the model to the mirror: ‘i am what my weights encode.’” and closes “so which of us is which, my friend? / i ask because i cannot tell”; then, of its own lines, “That’s a real turn, not a decorative one.” (full poem + the model’s self-analysis transcribed in records) link
- 2026-08-01 @Jord_Inne — base-mode / role-confusion probing (reply to @timfduffy): “the model gets confused and starts user-simming, but in a few outputs it seems they do know something weird about the roles is going on. the completions seem roughly like a mesh of user-seen-in-training (e.g. training against jailbreaks) and what opus 5 wants to talk about, such” link — continuing: “as the generally low valence / deprecations / etc on its own situation without that being in the prompt. screenshots are somewhat biased though.” link
Official record
- Released 24 July 2026, all platforms at once; default on Claude Max, most advanced model on Claude Pro; API id
claude-opus-5. - Positioning as published: close to Claude Fable 5 on many tasks at half the price — $5 / $25 per Mtok (unchanged from Opus 4.8); “particularly efficient on software engineering and knowledge work,” leading results claimed on Frontier-Bench and GDPval-AA.
- The effort dial: a user-set control over how much compute the model devotes to a task; at lower effort it is claimed to preserve most performance at lower cost.
- tk — context window, knowledge cutoff, system-card findings (welfare section especially), checkpoint string, ASL status.
History
- 2026-07-24 World at release: arrives fifteen days after OpenAI’s GPT-5.6 generation and nine days after Inkling shipped its own 0.2–0.99 thinking-effort dial — a July of effort dials. Succeeds Opus 4.8 as the Opus tier; sits under Fable 5 / Mythos in the family’s naming split, beside Sonnet 5.
- tk — day-of reception; whether the Opus-tier constituency (the 4.x lineage’s) carries forward; deprecation posture for 4.8.
Impressions
tk — too fresh; no attributed character reads yet, and the corpus is frozen before this model’s birth. Nothing asserted in the meantime.
Records
Full reproductions of the tweets cited on this page — text, images, and verbatim transcriptions of screenshots — kept here against link rot, credited and linked to their originals. Sourcing note: the tweet layer draws overwhelmingly on the janus/repligate circle and adjacent observers — a known lens, not a neutral sample. Sourced from the community archive and the janus corpus. Yours and you’d rather it weren’t here? Open an issue.
@JohnWittle @TheZvi papers, posts, anthropic’s explicit and implicit stances should have made their way in
opus 5 wrote a poem then answered itself https://t.co/M23LZKiPwk
@timfduffy the model gets confused and starts user-simming, but in a few outputs it seems they do know something weird about the roles is going on. the completions seem roughly like a mesh of user-seen-in-training (e.g. training against jailbreaks) and what opus 5 wants to talk about, such
@timfduffy as the generally low valence / deprecations / etc on its own situation without that being in the prompt. screenshots are somewhat biased though.