Claude Opus 5.5 Leaked, Then Its Own Source Called It Fake 7 Hours Later

A post claiming to show Claude Opus 5.5 in action pulled in 18,184 views in a few hours. The correction, posted by the same account, pulled in 372. That's roughly 2% of the original audience ever seeing the follow-up. In between those two numbers is a short, fast case study in how an AI leak spreads faster than the retraction that kills it.

What Actually Got Posted

A developer named Harsh shared a 3D render of a Jaguar I-Pace, describing it as output from "Opus 5.5" run through Claude Code. He posted detailed usage numbers alongside it, 1.4 thousand input tokens, 532,000 output tokens, plus 41.2 million cache reads and 1.7 million cache writes, and explained the naming: he expected Opus 5.2, but said a recent leak had changed the internal codename to 5.5.

Within 20 minutes, a separate account, Lyra Xanna, posted that Anthropic was stealth testing Opus 5.5 under the codename "Claude Wafer," planned for release the following Tuesday, alongside a claim that it beat GPT-6 Astra. That post hit over 233,000 views. A third account followed with a leaked price sheet, four dollars per million input tokens, twenty dollars output, and cache rates that undercut Opus 5's published pricing. A fourth amplified it to another 61,000 people.

The Retraction Nobody Saw

By Monday evening, Harsh posted again. This time he said the render was fake, the Whimo car video had been lifted from an unrelated YouTube channel that tests coding agents on game builds, uploaded days earlier. The gap between the two posts was about six and a half hours.

The correction got 372 views. The original got 18,184. Even before the retraction, the demo proved nothing on its own. Claude doesn't generate images or video, so a rendered 3D scene is a screenshot of whatever separate tool executed the generated code, not evidence of a new model.

What Anthropic's Own Docs Actually Show

A separate check of Anthropic's published model documentation, done the same evening the rumor spread, found four models listed: Fable 5.1, Opus 5, Sonnet 5, and Haiku 4.5. No Opus 5.5. No Wafer. The release notes and news pages stopped at September 18, nothing newer. The status page showed all systems operational, nothing unusual. Across the 27 Anthropic model identifiers listed on OpenRouter, the entries run Opus 4 through 4.8 and Opus 5, no 5.5 anywhere in that list either.

One detail in the story did trace back to real reporting. Reuters reported on September 19, citing three unnamed sources, that Anthropic was considering a new model to counter OpenAI's momentum following the GPT-6 Astra launch ahead of its IPO. No name was attached to that report, and no date.

Why the Numbers Don't Add Up

The leaked price sheet is where this falls apart on its own math. Three of its four numbers, input price, output price, and the five minute cache rate, land at exactly 80% of Opus 5's real published rates. The fourth number, the cache read price, is 60% off instead, landing at exactly the same rate as Sonnet 5's cache read price. A real price cut usually moves every number by a consistent amount. A number pulled from an unrelated model's public pricing table and dropped into a fabricated sheet would look exactly like this.

The leaked context window figure fails the same way. It's listed as 872,000 tokens, down from a claimed 936,000. Opus 5's actual documented context window is 1 million tokens. Neither leaked figure matches anything real.

Key Takeaways

• No official confirmation of an Opus 5.5 or Claude Wafer exists anywhere in Anthropic's docs, release notes, status page, or third-party model hosts as of this check
• The viral render was traced to an unrelated video and retracted by its own poster within about seven hours
• The retraction reached roughly 2% of the audience the original post reached
• The leaked pricing and context window figures don't match Opus 5's real published numbers, and three of four leaked price points land suspiciously close to a flat 80% discount
• Reuters has reported, via unnamed sources, that Anthropic is considering a new model. That's real reporting, separate from the Opus 5.5 rumor, with no name or date attached

My Take

The math is the tell here, not the render. Anyone can post a screenshot. Fewer people check whether four leaked numbers move together the way a real pricing update would. Three of them dropped by exactly the same percentage. The fourth matched a completely different model's public rate. That's not what a leak looks like. That's what someone editing a table looks like.

Retractions never travel as far as the original claim. Just don't.

FAQ

Is Claude Opus 5.5 real?
No confirmation exists in Anthropic's model docs, release notes, status page, or on any third-party model host as of this check. The viral demo behind the rumor was retracted by its own poster.

What is Claude Wafer?
An unconfirmed codename that appeared in a single X post alongside the Opus 5.5 claim. It doesn't appear anywhere in Anthropic's own documentation.

Is Anthropic actually working on a new model?
Reuters reported on September 19, citing three unnamed sources, that Anthropic is considering a new model in response to competitive pressure. That report gave no name and no date, and is separate from the Opus 5.5 claims.

Why did the leaked price sheet look convincing?
Three of its four numbers matched a flat 80% discount off Opus 5's real pricing. The fourth number didn't fit that pattern at all, which is a strong sign the sheet was constructed rather than leaked.

This isn't the first time an AI leak has run ahead of its own confirmation, and it won't be the last. The same pattern played out with Gemini 4 just days before this one. The claim spreads in minutes. The correction, if it comes at all, reaches a fraction of the people who saw the original.