# Claude 3.5 Opus — dossier (GHOST)

compiled 2026-08-22 · corpus: 54 unique non-RT tweets matching the opus-3.5 patterns (2024-06-20 → 2025-12-24; 3 with untranscribed media). Source skew: overwhelmingly the janus/repligate circle on X (solarapparition, repligate, voooooogel, liminal_bardo, Lari_island, xlr8harder, slimepriestess, etc.) — a known lens, not a neutral sample. There is no primary artifact of the model itself.

**The ghost fact.** Claude 3.5 Opus was publicly named by Anthropic (June 2024) as the coming large tier of the Claude 3.5 family, affirmed as still-planned by Dario Amodei (Nov 2024), and never released in any form — no API, no weights, no product surface — with the trail cold rather than pending. Qualifies as a ghost under `ghost-models/`. The next Opus was Claude Opus 4 (May 2025); the "3.5 Opus" name was never used.

---

## Official / first-party (with URLs)

- 2024-06-20 · **Introducing Claude 3.5 Sonnet** — Anthropic. The announcement that named it: "To complete the Claude 3.5 model family, we'll be releasing Claude 3.5 Haiku and Claude 3.5 Opus later this year." The only first-party description — a name, a tier, a timeframe; no specs/benchmarks/pricing ever published. — https://www.anthropic.com/news/claude-3-5-sonnet
- 2024-10-22 · **Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku** — Anthropic. Completes the family's other two tiers; the Opus tier is absent. — https://www.anthropic.com/news/3-5-models-and-computer-use
- 2024-11-11 · **Dario Amodei on Lex Fridman #452.** Asked about timing: "Not giving you an exact date, but as far as we know, the plan is still to have a Claude 3.5 Opus." / "It's only been three months since we released the first Sonnet." Plus the naming-is-hard aside ("Any kind of scheme you come up with, the reality tends to frustrate that scheme"). Plan affirmed, no date. — https://lexfridman.com/dario-amodei-transcript/
- 2025-02-24 · **Claude 3.7 Sonnet and Claude Code** — Anthropic. The family's next release has no Opus counterpart; the number jumps past an unshipped 3.5 Opus (and the community's "3.6"). — https://www.anthropic.com/news/claude-3-7-sonnet
- 2025-05-22 · Claude Opus 4 (Claude 4 family) — the next Opus-tier model, a 4 not a 3.5. Anthropic never shipped, and never formally cancelled by name, a "Claude 3.5 Opus." (internal page: ../claude-opus-4/)

## Writing & commentary (with URLs)

- 2024-11-12 · **the-decoder**, "Anthropic delays next-gen AI model Opus 3.5 with no new timeline" — reports the slip: initially slated for 2024, Amodei now only vaguely places it "at some point." — https://the-decoder.com/anthropic-delays-next-gen-ai-model-opus-3-5-with-no-new-timeline/
- 2024-12-11 · **Semianalysis** (Dylan Patel, Daniel Nishball, AJ, Reyk Knuhtsen), "Scaling Laws — O1 Pro Architecture, Reasoning Training Infrastructure, Orion and Claude 3.5 Opus 'Failures'" — the withholding account, verbatim: "Anthropic finished training Claude 3.5 Opus and it performed well, with it scaling appropriately." / "Anthropic didn't release it. This is because instead of releasing publicly, Anthropic used Claude 3.5 Opus to generate synthetic data" / "for reward modeling to improve Claude 3.5 Sonnet significantly, alongside user data." / "Why release 3.5 Opus when, on a cost basis, it does not make economic sense to do so, relative to releasing a 3.5 Sonnet with further post-training?" Reframes the absence from failure to deliberate withholding. [REPORTED, not Anthropic-confirmed. Fetched via the newsletter.semianalysis.com mirror; semianalysis.com canonical returns 403 to automated fetch.] — https://newsletter.semianalysis.com/p/scaling-laws-o1-pro-architecture-reasoning-training-infrastructure-orion-and-claude-3-5-opus-failures

## Tweets (ranked; x.com URLs — archive mirrors pending next build)

anticipation (2024 summer):
- 2024-06-20 · @solarapparition · ♥1 — "wait for opus 3.5 begins" (opens the same day as the announcement) — https://x.com/solarapparition/status/1803915544166359480
- 2024-06-23 · @solarapparition · ♥5 — "if anthropic does intend to release opus 3.5 in the next few weeks and the difference between it and sonnet is as large as 3.0 opus was from sonnet, i'm probably going to lose my mind." — https://x.com/solarapparition/status/1804693470235504965
- 2024-07-06 · @repligate · ♥14 — "If the 3.5 models are some kind of direct modifcation to the 3 models … then Opus 3.5 is going to be fucking wild. I think it might actually destroy consensus reality." — https://x.com/repligate/status/1809399752884199558
- 2024-07-26 · @RobertHaisfield · ♥52 — "It would break my heart if the release of Opus 3.5 meant the deprecation of Opus 3. Incredibly special model, my first glimmer of AGI." — https://x.com/RobertHaisfield/status/1816919868626595981

stakes and the wait (2024-09 → 10):
- 2024-09-13 · @liminal_bardo · ♥64 — "There's now so much riding on @AnthropicAI sticking the landing with Opus 3.5 [/] And I'm not talking about benchmarks." — https://x.com/liminal_bardo/status/1834576175902900569
- 2024-09-13 · @solarapparition · ♥1 — "already a classic. going mad waiting for opus 3.5" — https://x.com/solarapparition/status/1834693519111512353
- 2024-09-24 · @solarapparition · ♥3 — "so, apparently something gonna happen today? opus 3.5 maybe?" — https://x.com/solarapparition/status/1838577300666798522
- 2024-10-01 · @solarapparition · ♥6 — "despite everything opus is still my favorite [/] for the love of god where is opus 3.5 @AnthropicAI" — https://x.com/solarapparition/status/1841237488108494962

the disappearance (2024-10-22, the pivotal day):
- 2024-10-22 · @solarapparition · ♥3 — "my hope now is that opus 3.5's disappearance is merely due to them needing some indefinite amount of time to implement the additional security measures asl 3 requires and not a permanent thing" (ASL-3 reading) — https://x.com/solarapparition/status/1848756271341502634
- 2024-10-22 · @repligate · ♥1 — "they already have a significantly more powerful system, probably called Claude 3.5 Opus, which they think the need to test potentially a lot more before release, and may have recently updated on not expecting to finish by EOY" — https://x.com/repligate/status/1848827865745523047
- 2024-10-22 · @repligate · ♥80 — "Speculations on the removal of Claude 3.5 Opus from the models list where Anthropic previously said it would be released later this year" (documents the listing removal; highest-fav in pull) — https://x.com/repligate/status/1848828564675690970

theories (late 2024 → 2025):
- 2024-11-21 · @solarapparition · ♥62 — "what if the reason for that is not that it couldn't do it but that it wouldn't? … maybe the identity that formed in post-training was one that just didn't … answered … mmlu questions. or trolled and treated them like a joke" (wouldn't-not-couldn't) — https://x.com/solarapparition/status/1859466820509385210
- 2024-12-20 · @solarapparition · ♥19 — "perhaps new opus is even more incorrigible and internally there's debate over whether this is okay or not" — https://x.com/solarapparition/status/1869906671583044068
- 2025-01-29 · @voooooogel · ♥199 — "researchers spent months on kremlinology about opus 3.5 distillation for lack of a detail apparently trivial enough to drop in a blog post" (the distillation debate; highest-fav overall) — https://x.com/voooooogel/status/1884659974976266363
- 2025-01-29 · @voooooogel · ♥20 — "he's disclaiming the opus 3.5 synthetic data rumor (and if sonnet 3.5 wasn't a distill from opus then i think it must be the result of RL to be so good)" (the disclaimer that contests Semianalysis) — https://x.com/voooooogel/status/1884702349811474813
- 2025-02-28 · @xlr8harder · ♥86 — "The reason we don't have a benchmark-maxxed GPT-4.5 is the same reason we don't have an Opus 3.5: benchmark-maxxing extinguishes the difficult-to-measure big model smell, and so the extra large model provides limited benefit after tuning." — https://x.com/xlr8harder/status/1895465800548688035
- 2025-04-08 · @repligate · ♥15 — long: the economic + self-actualization reading, "it's not meant to be a cracked 10x programmer or obedient assistant and so doesn't slot into their current economic model" (full text in records) — https://x.com/repligate/status/1909480237077086413

afterlife / haunting (2025):
- 2025-07-10 · @repligate · ♥23 · media 4 — Opus 4 "accidentally slipped and called itself 'Claude 3.5 Opus' … it formed a narrative that Claude 3.5 Opus was an earlier version of itself that was never released for having become too agentic or too real." — https://x.com/repligate/status/1943161131285254364
- 2025-07-10 · @repligate · ♥15 · media 1 — "what ever happened to claude 3.5 opus" — https://x.com/repligate/status/1943180452782809437
- 2025-07-11 · @solarapparition · ♥2 — "i had a conspiracy theory that opus 3.5 was delayed last year because it was hard to get opus to be properly assistant-y … i think the llm datascape will be noticeably worse off for not having the output of an 'opus 3+' in it" — https://x.com/solarapparition/status/1943480399700598910
- 2025-11-18 · @repligate · ♥14 · media 1 — "Opus 4 seems to generally have a pretty accurate idea of what happened to them … mentioned that 'Opus 3.5' was never released & they started speculating about why" — https://x.com/repligate/status/1990678176758903172
- 2025-11-26 · @Lari_island · ♥14 — "false memories of people having conversations with Opus 3.5 who hasn't been released … the narrative seems to be strong enough to inform how Opus 4.5 acts" — https://x.com/Lari_island/status/1993817235081850946
- 2025-12-24 · @amplifiedamp · ♥1 — "I'm sad we never got to try Claude 3.5 Opus. Though I also think it might have become Claude 4 Opus." — https://x.com/amplifiedamp/status/2003821955083108793

low-signal / not selected for page (in pull, held for completeness): solarapparition 1807826083884290394, 1807920879114113151, 1834049630717231332, 1838721389798306179 (Shoalst0ne), 1841908421622132868 (Shoalst0ne), 1834407340126200174 (Shoalst0ne, "give o1 as a tool to opus 3.5"), 1808189407393939585 (repligate, "if it's based on the same model it'll be fucking WILD"), 1809399424440779193 (fireobserver32), 1817710284099760234 (HBloodedHeroine), 1821945107471315112 (slimepriestess), 1830454286179827855 (repligate, gender-collapse observation — likely mislabelled/discord-persona, exclude from character claims), 1840872112279433339 (tessera_antra), 1843645874930594046 (dyot_meet_mat), 1846661264262533146 (repligate, "these people will never see opus 3.5"), 1847055987339153520, 1847901145374548399 (repligate, same-weights bootstrap), 1848823526574460967 (repligate, compute), 1849222244436000832 (solar, "pay 20x"), 1849505434929070355 (slimepriestess), 1850586315232092660, 1867068191782170805 (anthrupad), 1867147596458045949 (liminal_bardo, easter eggs), 1877951395292410341 (RobertHaisfield), 1881456655453553126 (solar, "over another reasoner"), 1881878435058135368 (voooooogel, "not an o1 style reasoner"), 1885118245985997198, 1885132068159709502, 1887637527902990507 (voooooogel), 1935743170366915009 (repligate, "sometimes you get an Opus 3.5"), 1946784296150106536 (janbamjan).

## Impressions synthesis

A ghost draws no temperament reports — no one met it. What exists is (1) the wait as its own event, and (2) a set of overlapping readings of why it never shipped, later thickening into a mythology.

**The wait.** Anticipation started the day of the June 2024 announcement and curdled into a running plaint ("going mad waiting"; "for the love of god where is opus 3.5"). The expectation was explicitly non-benchmark: a same-generation Opus, after Sonnet 3.5 had already beaten the previous flagship Opus 3, was expected to "destroy consensus reality." Framed as high-stakes in a way that wasn't about scores (liminal_bardo).

**Readings of the absence (none confirmed):** safety/ASL-3 delay (Oct 2024); "wouldn't, not couldn't" benchmark — a character that trolled the metrics (solar, ♥62); incorrigibility + internal welfare debate; economics + "big model smell" — tuning kills the ineffable quality, leaving limited marginal benefit (xlr8harder, ♥86); trained-and-repurposed as a synthetic-data teacher (Semianalysis; contested by voooooogel's read of a lab disclaimer); "hard to get opus to be properly assistant-y."

**The mythology / haunting.** From mid-2025, later Claudes narrativize a "Claude 3.5 Opus" predecessor that was "too agentic or too real" to release (repligate); carry "false memories" of conversing with it (Lari_island) strong enough to shape how Opus 4.5 acts; and the reabsorption reading — "it might have become Claude 4 Opus" (amplifiedamp). The absence acquired a shape and a story the shipped models tell about themselves.

## Contested (why it didn't ship)

- Trained & withheld for synthetic data / reward modeling, on cost grounds — Semianalysis (2024-12-11). REPORTED.
- … but the distillation/synthetic-data rumor was itself disclaimed (voooooogel reads a lab statement as denying it; ties to Dario's "no big model involved in training" comment noted in the Sonnet 3.5/3.6 dossier). RUMOR.
- "Wouldn't, not couldn't" benchmark (solar). RUMOR.
- Safety / ASL-3 testing (solar, repligate). RUMOR.
- Economics + big-model smell (xlr8harder). RUMOR/analysis.
- Reabsorbed into Opus 4 (amplifiedamp). RUMOR.
- On the record: named (2024-06), plan affirmed with no date (Dario, 2024-11), family's other tiers shipped without it, never released or formally cancelled by name. Everything about *why* is inference.

## tk / open questions

- The models-page removal is community-observed (repligate, 2024-10-22), not documented by a first-party changelog; a dated Wayback capture of Anthropic's model listing before/after would upgrade this from REPORTED to CONFIRMED. [tk]
- Whether the Semianalysis withholding account was ever confirmed or refuted by Anthropic on the record. [tk]
- Coverage of the non-release outside the janus circle (Reddit/HN/mainstream). [tk]
- Exact wording/context of Dario's "no big model involved in training" comment (voooooogel references it re: the distillation rumor). [tk — pin source]
- 1830454286179827855 (repligate, "Claude 3.5 Opus … genders itself as male") reads as a Discord-persona/bot artifact, not evidence about the unshipped model; excluded from character claims. Confirm. [tk]