Claude Opus 5.5 Leak: A Checkpoint Called "claude-wafer-eap" and a Rumored Tuesday Release

The claim doing the rounds this weekend is specific enough to check and thin enough to fall apart under the check: a new checkpoint for the next Opus update, now labelled Claude Opus 5.5, was reportedly already in hand, with an internal codename — "claude-wafer-eap" — and a release that could land as early as Tuesday, September 22. The source is the X account @kimmonismus relaying a leaker called Lyra. There was no model card, no API identifier, no price sheet, no benchmark, and no second outlet reporting the same checkpoint string — the date was the only part of it with anything behind it. The model shipped anyway, on Tuesday, September 22, and it is callable today; the codename and the leaker handle behind it are still uncorroborated. What there was, underneath the claim, is a real and well-documented event: Claude Opus 5.2, the unreleased Opus build that was being served to a slice of traffic under an unchanged "Opus 5" label through mid-September, and which the vendor briefly cut off on the night of September 17. That much is corroborated across multiple outlets. Everything above it — the checkpoint, the number, the codename — was one account's word. And there is a third thing worth noticing, which is what the leak does to the version number itself: Claude Opus 5 shipped on July 24 at $5.00/$25.00 per million tokens and led the tier until September 22; Claude Fable 5.1 arrived September 1. The next Opus is 5.5, not 5.2, so the naming has jumped past 5.3 and 5.4 — and the number has done more work in this story than any benchmark would.

What the leak actually claims

Strip the packaging and there are four assertions. First, that a checkpoint exists — a model that has been trained, saved and is being staged somewhere. Second, that its version number is 5.5, not 5.2 or 5.3. Third, that it carries a codename, "claude-wafer-eap", with "eap" reading as early-access program in the usual vendor shorthand. Fourth, that release is imminent, potentially Tuesday.

Each of those is a different kind of claim, and they are not equally supported. The fourth is the one with external evidence behind it, and that evidence points at a date, not at a codename. Prediction markets on the next Claude Opus release opened on September 16 and had clustered the bulk of their implied probability across September 21 to 24, with a meaningful share of volume still backing "no release by September 30." The market's resolution rules are narrow: the model has to be publicly accessible, and it has to be explicitly named an Opus. Fable, Mythos, Sonnet and Haiku releases do not count. "Claude Opus 5.5" appears in that market's own list of qualifying names — which is where the string most plausibly entered circulation, and which is a list of possibilities, not a report of a sighting.

The codename is the weakest link. "claude-wafer-eap" appears in the signal text and, as far as we can establish, nowhere else: no repository reference, no API slug, no configuration leak, no corroborating account. Codenames are exactly the kind of detail that spreads because it sounds verifiable. Treat this one as unverified until a second, independent source produces the same string. The same applies to the "Lyra" attribution — a leaker handle with no track record we can check against.

What is actually verified

The verified baseline is Claude Opus 5, and it is unusually well documented, which is what makes it useful as a yardstick. Anthropic shipped it on July 24, 2026 — no 5.1 in between — with a 1M-token context window, up to 128K output tokens (300K via a beta header on the Batch API), adaptive thinking on by default, and five reasoning effort levels from low to max. Pricing is flat at $5.00 per million input tokens and $25.00 per million output tokens at every prompt length, the same as Opus 4.8 before it, with the Batch API at half that. Artificial Analysis placed it near the top of its Intelligence Index, with the exact score depending on which effort variant you look at — a detail worth holding onto, because a "5.5" scoreboard will not be comparable to a "5" scoreboard unless the effort setting matches.

On top of that sits the corroborated September event. From roughly September 14 to 15, developers reported that a small, probabilistic share of requests nominally addressed to Opus 5 were being answered by a different backend model, with request metadata pointing at an unreleased claude-opus-5.2 — a version number that skips 5.1 entirely. Reported behaviours included terser output and unprompted self-iteration: writing code, running tests, fixing failures, repeating until the tests pass, rather than stopping half-finished. The detection trick that circulated alongside it — asking the model a specific question about a person it should not know, with search and memory disabled — was a way of telling which backend answered. On the night of September 17 Anthropic cut the channel without notice, an event the developer community dubbed the "midnight shock"; the channel came back about a day later and, by those reports, widened from Claude Code to the chat and cowork surfaces. No model card, no pricing table, no API identifier, and no official statement accompanied any of it.

The version number is the story

Here is the structural point the leak was circling, and it is the part that mattered. Anthropic has already demonstrated, twice this year, that it will iterate the Opus tier without a launch event — 4.8 in late May, 5 on July 24 with 5.1 skipped, then a 5.2 that existed in production without existing in public. A vendor that ships through backend routing and controls the rollout ratio can also decide, at the last moment, what number to print on the box. It was announced as 5.5 rather than 5.2, which is a marketing decision made after the technical work, and it is the kind of decision a leaker hears about late and imprecisely.

That cuts both ways for anyone reading this. A leaked version number is weak evidence about capability and strong evidence about positioning. The number really is 5.5, which implies a larger step than 5.2 would have — and it implies Anthropic wants the market to read it that way, ahead of an IPO the company has reportedly pushed toward November. Reuters reported on September 18, citing three sources, that Anthropic is weighing whether to ship a new model specifically to answer GPT-6 Astra's enterprise momentum, while its chief executive publicly asks the industry to slow down. A bigger number on the same checkpoint costs nothing and buys a headline. That is not a reason to disbelieve the leak; it is a reason not to price the number into a decision.

What would have to be true on Tuesday

A release on September 22 was checkable in about five minutes, and naming the checks in advance is what kept the announcement from being absorbed uncritically. They are the same five worth running against any new Opus.

• Availability — a public model card, an API model identifier that returns completions, or an open rolling waitlist. A closed beta does not qualify, and prediction markets generally resolve against it. This is the line that separates "shipped" from "tested on some accounts."

• Pricing — whether the new Opus holds the $5.00/$25.00 tier, moves it, or arrives with a promotional rate. Opus 5 held Opus 4.8's price exactly, which was the pattern to check — and the one that broke: Opus 5.5 came in below Opus 5 on every line.

• Context and output limits — Opus 5's 1M context and 128K output (300K on Batch) are the current floor. A new Opus that ships with less would be a downgrade dressed as a launch.

• Independent scoring — an Artificial Analysis entry at a stated effort level, not a vendor chart. One now exists: Artificial Analysis carries Claude Opus 5.5 at 58 on its Intelligence Index, in the configuration it labels "Adaptive Reasoning, Max Effort, Default Fallback" (read from its model page on September 24, 2026). The setting is the point — a max-effort run is not interchangeable with a default-effort baseline, and a vendor table, run at whatever effort the vendor chose, is not a substitute for either.

• The routing behaviour — whether the terse, self-iterating behaviour reported in the gray-test accounts is still present in the public build, or whether it was a testing artefact. That behavioural change is the most concrete thing developers actually observed, and it is the thing most likely to be quietly tuned out before the public build.

Four of the five resolved on September 22. The API identifier is claude-opus-5-5 and it returns completions. The tier moved down rather than up: $4.00 per million input tokens and $20.00 per million output, with cache reads at $0.20, all of it Anthropic's published rate rather than a promotional one. Context and output hold at 1M and 128K, and independent scoring followed. The routing behaviour is the one check a model card cannot answer, and it is the one still open: whether the terse, self-iterating style reported in the gray-test accounts survived into the public build.

Trying an unproven Claude without betting a production path on it

The awkward gap in a week like this one is that the model everyone wants to evaluate is the one nobody can call. That gap closed on September 22, which leaves the ordinary problem that comes with any new frontier model: how to point a slice of real traffic at it without rewriting your integration or committing a critical path to weights nobody outside the vendor has stress-tested on your workload.

That is the case routing is built for. OrcaRouter puts the current Claude lineup behind one API key at provider list price with 0% markup — the newest entry is the one this article is about: anthropic/claude-opus-5.5 at Anthropic's published list price, $4.00 per million input tokens and $20.00 per million output, with cache reads at $0.20 and a 1M-token context, passed through with nothing added, so a vendor price change is live on our side the same day rather than after a sync. The rest of the Claude lineup is on the same key at list too: Claude Opus 5 at $5.00/$25.00, Claude Fable 5.1 at $10.00/$50.00, Claude Opus 4.8 at $5.00/$25.00 and Claude Sonnet 5 at $2.00/$10.00. Claude Opus 5.5 is one more route rather than a second contract: you can send a percentage of traffic at it, keep automatic failover pointed at a model you have already characterised, and compare outputs on your own prompts instead of on a vendor's chart. Composing a call across several models — a draft from one, a verification pass from another — is a routing-DSL line rather than a new service.

To be explicit about what we are and are not saying: Claude Opus 5.5 is one of our routes. anthropic/claude-opus-5.5 is live on OrcaRouter at Anthropic's published list price with 0% markup — $4.00 per million input tokens, $20.00 per million output, $0.20 per million cache reads, 1M-token context — verified against our own model list on September 24, 2026. What this article is not is a product page, and what the leak analysis above is not is independent evidence: the codename still has no second source, and the vendor's benchmark table is the vendor's, run at the effort settings the vendor chose.

What to watch, and what to ignore

The model identifier was the artefact that settled it, and it is the one a leak cannot fake for long: the Claude API lists claude-opus-5-5, and from there it is reachable. The pricing table came with it, and it said something the leak did not — Opus 5.5 sits a tier below Opus 5, at $4.00/$20.00 with cache reads at $0.20 rather than $0.50. The gray-test channel told the truth as well: cut on September 17, restored about a day later, with the public model arriving within the week.

Ignore, for now, the codename. "claude-wafer-eap" is a single-source string with no corroboration, and the same goes for the "Lyra" attribution behind it. Ignore the version number as evidence of anything except positioning. And treat any scoreboard that appears in the first 48 hours of a release with suspicion — including ours — until there is something real on both sides of the comparison.

The leak's date and its version number both turned out to be right, and it is worth being precise about what that does and does not mean. They were right as predictions, not as reporting: the number had already appeared in a prediction market's own list of qualifying names before anyone claimed a sighting, and the codename — the detail that made the claim feel specific — has still never been produced by a second source. The September Opus 5.2 event is the durable lesson, because it will repeat: a model can be served in production for days before anyone can name it, and the number finally printed on the box is a positioning decision made after the technical work is done.