Models·8 min read
By BitsMindsSource: Anthropic model documentation

Claude Opus 5.5: One Rumour, Three Names in 16 Days

Anthropic shipped Claude Opus 5.5 on 22 September at $4 and $20 per million tokens — the price and the day the leak named. The sixteen days before it were a different story: one unconfirmed release, three different names, and a chain that had already been wrong about two of them.

Claude: three names, one unconfirmed rumour A warm paper evidence folder bears the official Claude mark. Three removable labels read Fable 5.2, Opus 5.2 and Opus 5.5 with a question mark. Beside them, a large brass and charcoal magnifying glass reveals another question mark on a blank specimen card. The illustration represents changing names and missing confirmation, not an announced model. Original vector editorial illustration for https://www.bitsminds.com/news/claude-opus-5-5-rumour-three-names. Article context: 21 September 2026. Model names are rumoured. Claude mark uses the unchanged outline from the project logo asset. Fable 5.2 Opus 5.2 Opus 5.5? BITSMINDS.COM
Share:

Update, 22 September 2026 — it shipped, and the leak was right. Anthropic announced Claude Opus 5.5 on the day the leak named, at the prices it named: $4 per million input tokens and $20 per million output, with cache reads at $0.20 and cache writes at $5. The model ID is claude-opus-5-5, the context window is 1M tokens with 128K of output, and it is live on the Claude API, Claude Code, AWS, Google Cloud and Azure. Anthropic says it "performs at the level of Claude Fable 5.1 on most work" while costing 40% less to run than Opus 5 — note that the sticker fell 20% but the cost of a task fell 40%, because it also spends about 20% fewer tokens getting there. One more leaked detail held: thinking is now always adaptive, and switching it off is no longer an option. Everything below was written on 21 September and is left exactly as it stood. Our full launch coverage has the pricing, the benchmark charts and what changed in how the model thinks.

As of 21 September 2026, Anthropic has not announced a Claude Opus 5.5. The company's model table lists the same four current models it listed a week ago — Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5 — and the newest model post in its newsroom is still the Fable 5.1 and Mythos 5.1 launch of 1 September. There is no model page, no API identifier and no pricing row for an Opus 5.5. Opus 5 is still $5 per million input tokens and $25 per million output.

What has changed is the rumour. The same unconfirmed release has now been given three different names in sixteen days, and each rename has arrived with more specific detail than the last.

Sixteen days, three names, one modelWhat the leak chain published, against what Anthropic published · September 2026Rumoured launch: 22 Sep1 Sep5 Sep9 Sep13 Sep17 Sep21 SepLEAKDOCSFable 5.2One X post. No name given.Opus 5.2“Felt different” + a trivia probe.Opus 5.5A codename and a full price sheet.Fable 5.1 and Mythos 5.1Anthropic's last model announcement — nothing since
Each iteration of the rumour has arrived with more specific detail and no more evidence. Dates are when the claim was first posted, not when aggregators ran it. Sources: Anthropic model documentation and newsroom for the confirmed track; leak dates as reported by the trackers cited in this article.

The first version surfaced on 5 September, when the X account @kimmonismus wrote that a major Anthropic model was close and hedged the name itself — "Fable 5.2, or whatever it ends up being called". That prediction's first date, 9 September, passed with nothing. We covered that version on 14 September, and concluded then that the leak was thin but the release arithmetic was not: Anthropic shipped six frontier models between April and September at a median gap of 24 days, which puts the next one due around now.

The second version appeared in the middle of the month, when developer posts claimed an unreleased Opus 5.2 was being served to some traffic under the Opus 5 label. The evidence offered was that responses "felt different" in Claude Code, plus a trivia probe — a question about a recent news subject that the new checkpoint could supposedly answer and Opus 5 could not. Trackers disagree about even this much: OrcaRouter treats a 14–17 September serving window as having happened, while CellCog, which has tracked the same chain since 14 September, records that no API string, router catalogue entry or screenshot of a model list ever carried the name.

The third version landed on 20 September, when, in CellCog's phrase, the rumour renamed itself. The model is now Opus 5.5, with an internal codename of claude-wafer-eap — the suffix reading as early-access program — attributed to the X account @lyraxana and relayed by @kimmonismus. OrcaRouter, which went looking for the string, reports that it appears nowhere else: no repository reference, no API slug, no configuration leak, and no second outlet reporting the same checkpoint.

What is claude-wafer-eap?

claude-wafer-eap is a string, and for now that is all it is. It surfaced on 20 September in a post by the X account @lyraxana, was relayed by @kimmonismus, and has been reproduced by every write-up since without anyone producing the thing it is supposed to name. Read literally it parses as a wafer-codenamed checkpoint exposed through an early-access programme; the -eap suffix is the only part of it that follows a convention anyone can point to.

What makes the string checkable is that Anthropic model identifiers are public by design — every model the company sells answers on the Models API and gets a row in the documentation the moment it exists. OrcaRouter went looking for claude-wafer-eap and reports that it appears nowhere else: no repository reference, no API slug, no configuration leak, no second outlet reporting the same checkpoint. We checked Anthropic's public model list again on 22 September, the rumoured launch date itself. It lists Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5. There is no wafer, no Opus 5.5 and no Fable 5.2 in it.

The renames are the reason to hold the string lightly. The same chain called this model Fable 5.2 on 5 September and Opus 5.2 on 14 September, and CellCog records that no API string, router catalogue entry or screenshot of a model list ever carried either name. claude-wafer-eap arrived after two names had already failed, from a single account, with no second detail attached that anyone can test — weaker evidence than a codename that arrives cold, which is what "Honeycomb" was three days before Opus 5.

The price sheet is the new part

Every previous turn of this rumour dealt in vibes. This one comes with numbers. As reported by WinCentral and by 36Kr, the leaked sheet puts Opus 5.5 at $4 per million input tokens and $20 per million output, with cache reads at $0.20 and cache writes at $5 — a flat 20% under Opus 5 on both headline rates. The claimed launch date is Tuesday 22 September.

That combination is worth noticing precisely because it is checkable. A price cut on a named date is not the kind of claim that can be quietly walked back: either a pricing row appears or it does not. It is also the sort of detail that makes a rumour feel more solid without making it better sourced. The sheet has the same provenance as the codename, which is one account with no track record anyone has been able to establish.

The version number itself is the other oddity. Anthropic has been meticulous about decimals this year — 4.6, 4.7, 4.8, 5, 5.1 — and a jump from Opus 5 straight to Opus 5.5, skipping 5.2 through 5.4 after a fortnight of people insisting the next one was 5.2, is a strange thing for a company to do and an easy thing for a rumour to do.

The demos deserve the most scepticism

The loudest part of the current cycle is a run of 3D demonstrations attributed to testers with early access: a detailed Waymo self-driving car complete with Jaguar badge and lidar housing, reportedly generated from 1,400 input tokens and 532,000 output tokens; a browser-based car viewer built in Three.js; an interactive WALL-E console; a cliff-side villa with exploded views. 36Kr names the people behind several of them.

Read that list against July. Prompt-to-3D in a single file is exactly what Claude Opus 5 became known for on release — The Decoder covered it at the time, and the demos that circulated then were a playable Rocket League clone, a Three.js Lord of the Rings world and a space RTS. A new model that is better at the thing the current model is already famous for is a plausible claim. It is also the single easiest claim to manufacture, because the artefacts are indistinguishable from what the shipping model produces, and because a screenshot carries no model ID.

The aggregator layer is not helping. KuCoin's write-up, published the same day, credits the 3D improvements to "Fable 5.2" in one paragraph and to Opus 5.5 in another — inside a single article about a single model. When a chain of reports cannot hold the name of the thing steady, the details travelling along that chain have not been checked either.

What would actually settle it

The same three things as in July, and they are all public: a model ID that answers on the Claude API or turns up in a router's catalogue, a model page in Anthropic's docs, or a pricing line. The Models API will list a new model the moment it exists, which is why every real Anthropic leak so far has been anchored to a string rather than to a feeling. The "Honeycomb" leak that preceded Opus 5 by three days carried a codename and a context-window figure, and both turned out to be right.

None of which means nothing is coming. Anthropic's own retirement commitments — published for every model it sells, and readable backwards as a launch calendar — show a company that has shipped roughly monthly since April, and it has been three weeks since Fable 5.1. A release in the next fortnight would surprise nobody.

The gap between that and what is circulating is the whole point. "Anthropic ships often and is probably due" is a claim about a company's habits. "Opus 5.5 launches Tuesday at $4 and $20" is a claim about a specific artefact on a specific date, and it is being carried by the same single thread that has already been wrong about the name twice. If Tuesday passes without a pricing row, expect the fourth name by the weekend.

Note, 22 September 2026. The paragraph above was wrong about the outcome. There was no fourth name: the same thread that had already renamed the model twice got the third name, the date and the entire price sheet right, down to the cache rates. What it never produced was a checkable artefact — and the artefacts this article asked for are exactly what settled it. A model ID, a documentation page and a pricing row appeared on 22 September, and not one day before.

More on Claude

Evergreen coverage we keep current — start here.

Want AI news before everyone else?

The morning's most important AI stories, straight to your inbox. No fluff.

Related Articles

Kimi K3.1 Pro open weights Editorial illustration on a deep ink-blue field: a heavy machined case stands open with three glowing weight plates lifting out of it, and the Kimi logo is inlaid in the case front. The type column reads Moonshot AI open weights, Kimi K3.1 Pro, 2.8T weights, free. An editorial metaphor for the model's weights being published on Hugging Face; the license is Moonshot's own and carries conditions. 2.8T · MXFP4 MOONSHOT AI · OPEN WEIGHTS Kimi K3.1 Pro 2.8T weights, free BITSMINDS.COM
Models

Kimi K3.1 Pro Weights Go Live: a 2.8T Open Model

Mistral Large 4 weights released Editorial illustration on a deep red-orange field: an open wooden shipping crate with the Mistral logo stencilled on its side and warm light rising out of it, its lid leaning against it. The words open weights and Large 4 sit in the type column. The crate is an editorial metaphor for the weights becoming downloadable, not a Mistral product image. OPEN WEIGHTS Large 4 1 TRILLION PARAMS BITSMINDS.COM
Models

Mistral Large 4 Goes Open Weight, With a Hosting Catch

Google launches Gemini 4 Carbon Editorial illustration: a brilliant-cut diamond, carbon in its hardest form, catches the Gemini colour spectrum in its facets, with the Gemini logo inlaid in its lower half. Beside it the word Carbon. The diamond is an editorial metaphor for the model's name, not a Google product. GOOGLE DEEPMIND · GEMINI 4 Carbon FRONTIER MODEL BITSMINDS.COM
Models

Gemini 4 Carbon: Google’s New No. 1, Cyber Defenders First