Is Claude Opus 5.5 Out Yet? Every Claim, Checked
Claude Opus 5.5 is not out. As of September 22, 2026, Anthropic has announced nothing, its official model list still ends at Opus 5 and Fable 5.1, and every number circulating about the price, the context window and the benchmark position traces back to a screenshot of a post on X.
Key takeaways
- Nothing is announced. Anthropic's model overview lists Fable 5.1, Mythos 5.1, Opus 5, Sonnet 5 and Haiku 4.5, plus legacy models. There is no Opus 5.1 row, no 5.2 row and no 5.5 row, and the API release notes for September say nothing about a new Opus.
- The codename doing the rounds,
claude-wafer-eap, does not appear in any verifiable artifact. It is not in a model picker, not on OpenRouter, not on a leaderboard, and not in an API error string. That makes it weaker evidence than July's Honeycomb leak, which at least surfaced inside a shipping product's UI. - The leaked price list ($4 input and $20 output per million tokens) has no source attached to it. What is documented is that Opus 5 costs $5 input and $25 output, so the rumor is a 20% discount off a real number, which is exactly what a plausible invention looks like.
- The prediction market figure being quoted as "88% for today" was 81.8% on the Polymarket bucket for September 22 when we checked it. Either way it is a crowd's opinion priced in dollars, not a leak.
- The same codename-watching crowd called Fable 5.1's September 1 launch correctly one day ahead, and got the model name wrong in the same breath: the "Opus 5.1" it predicted alongside has still never shipped. Expect the week to be roughly right and the version number to be wrong.
- Anthropic's own cadence is the strongest real signal. Opus 4.7 to 4.8 was about six weeks, 4.8 to Opus 5 about eight, and Opus 5 shipped on July 24, 2026. On base rate alone something Opus-tier in late September or October is plausible. A base rate is not a schedule.
- The outage people are reading as a launch signal was real and is not one. Anthropic's status page logs elevated errors from 00:50 to 02:10 UTC on September 22 affecting Opus 5, Fable 5, Fable 5.1, Mythos 5 and Mythos 5.1. Serving incidents happen on ordinary days.
Last verified: September 22, 2026. This is a developing story built on unverified reports. Every price, context length and model ID presented below as fact comes from Anthropic's own model overview, newsroom and status page. Everything attributed to a post on X is labelled as such. We will update this page when Anthropic makes an announcement, or when any claim here is confirmed or disproven.
What is actually confirmed right now
Start with the boring part, because it is the part that is true. Anthropic's public model list, the one the API and the docs are generated from, currently looks like this. If a Claude Opus 5.5 existed in any shipping form, it would have a row here and a model ID you could send a request to.
Two things in that table matter for reading the rumor. First, Opus is not the top of the range any more: Fable 5.1 sits above it at double the price, which is the tier the leaks keep comparing a hypothetical Opus 5.5 against. Second, Opus 5 already has a 1M token context window, which quietly deflates one of the headline claims before we even get to it.
Anthropic's September newsroom output is also a useful negative signal. The posts that exist cover the Fable 5.1 and Mythos 5.1 launch on September 1, a research piece on measuring the pace of AI, a threat intelligence report, an Accenture partnership and a life sciences program. A frontier model launch is not something Anthropic ships quietly.
The codename: why claude-wafer-eap is weak evidence
The load-bearing claim in this cycle is that Anthropic is stealth-testing a model with the internal codename claude-wafer-eap, which will ship as claude-opus-5-5. The claim originated in posts on X and spread through aggregator accounts within hours.
We went looking for the string itself, because a real stealth test usually leaves a mark somewhere. It is not in OpenRouter's provider or model listings. It is not on any public leaderboard. There is no API error message, no changelog entry, no docs diff and no JavaScript bundle containing it. The word "wafer" does appear on OpenRouter attached to an unrelated third-party provider, which is the sort of coincidence that keeps a rumor alive for an extra day.
That absence is the whole story. Compare it with the leak we covered in July, when a model called Claude Honeycomb EAP appeared in Cursor's production model picker with real metadata attached and was pulled within hours. That was a screenshot of a shipping product's UI: something a third party had rendered, which is hard to fake and easy to corroborate. The current cycle has produced a screenshot of a claim.
The evidence ladder, top to bottom
- A string in a first-party artifact: docs, API response, model list, billing page.
- A string in a third-party product that integrates the API: a model picker, a router, a leaderboard.
- A named person with a track record describing something they saw, with a timestamp.
- A screenshot of an anonymous post asserting a fact.
- A prediction market price, which aggregates everything above including the noise.
Every claim in the Opus 5.5 cycle currently sits on rungs four and five. In July, the Honeycomb story reached rung two.
The leaked pricing, against what Opus 5 actually costs
The pricing screenshot is the most repeated part of the rumor, and it is worth laying next to the documented numbers, because the shape of the claim tells you something about its likely origin.
Read the middle column as a set and a pattern shows up: every figure is a clean 20% cut off a real, published Opus 5 number. That is not evidence of a leak. It is what you would produce in thirty seconds if you wanted a credible-looking price list, and it is also, to be fair, roughly what a genuine generational price cut tends to look like. The claim is unfalsifiable in either direction until an announcement lands, which is exactly why it travels so well.
One number in the circulating list we could not check at all. A cache write figure of $5 against a claimed Opus 5 baseline of $6.25 is being quoted, and the models overview page does not carry cache write pricing. We are not printing it as a comparison because we could not confirm the baseline, let alone the claim.
The context window claim is not a claim
"Context window reportedly around 900K to 1M" reads like a spec leak and is not one. Opus 5, Sonnet 5, Fable 5.1 and Mythos 5.1 all already document a 1M token context window. A new Opus with 1M context would be holding position, not advancing it. If anything, a range that tops out at the current number is a small mark against the leak's credibility: a real internal spec sheet would say 1M, or it would say something new.
The number that would actually be news is max output tokens, which sits at 128K across the current frontier models, or a change to how long-context requests are priced. Nobody is leaking those, which is usually a sign that nobody is looking at a real spec sheet.
The performance claim, and why the tier structure makes it interesting
The claim is that the new model is "way better than current Astra and Fable 5.1" while costing less than half of Fable. Take the benchmark part first: there is no benchmark. No chart, no eval harness, no named test, no third party reproducing anything. It is an assertion.
The pricing part is more interesting, because of where Opus now sits. Anthropic's range currently runs Haiku, Sonnet, Opus, then Fable and Mythos at the top. Opus at $5 and $25 is the workhorse frontier tier; Fable 5.1 at $10 and $50 is the premium one. A rumor that the next Opus matches the premium tier at a fifth of its output price is, structurally, a rumor that Anthropic is about to collapse its own price ladder. That is a big claim to make on the strength of an anonymous screenshot, and it is worth noticing that this is precisely the claim most likely to generate engagement.
Anyone who wants the current, measured picture rather than the rumored one has it already. We ran the three-way comparison of Opus 5, Fable 5 and the GPT-5.6 generation, and the lesson from that piece is worth repeating here: read the vendor's own charts directly, because secondary benchmark numbers get mangled in transit almost every time.
What the Polymarket odds actually say
A prediction market is the most cited and least understood input in this cycle. The figure being passed around is "88% for today". When we checked the Polymarket event on the next Claude Opus release, the leading date bucket was September 22 at 81.8%, on a market with roughly $265,000 of volume.
The resolution criteria matter more than the percentage. For any date in that market to resolve Yes, the model has to be explicitly named Opus and it has to be publicly accessible, which excludes a closed beta or a limited early access program. So the market is not pricing "will something ship", it is pricing "will a publicly available model with Opus in its name ship on this date", and a stealth test under a codename resolves nothing.
The deeper problem is circularity. The market moves on the same X posts that people then justify by pointing at the market. We have built trading automation against Polymarket, and the lesson from that work applies here: a market price is a useful summary of what a crowd believes and a terrible substitute for a primary source. One trader on X put it more bluntly, telling the forecasters to stop basing release-date predictions on Polymarket odds because it had become obvious.
Who said what, and when
The rumor has a traceable shape, and laying it out is the fastest way to see how thin the base is.
Notice that the chain has one origin and four amplifiers. TestingCatalog, which is a real outlet and one of the more careful ones in this niche, is transparent about it: its own piece rests on the social posts and says so. The coverage volume is not independent corroboration, it is the same claim with more retweets.
The precedent: right about the week, wrong about the name
This crowd is not useless. It has a track record, and the track record has a very specific shape.
On August 31, 2026, the same codename-watching accounts called the next day's launch correctly, naming two internal codenames with the same -eap suffix pattern. Anthropic shipped Fable 5.1 and Mythos 5.1 on September 1. The date was right, one day ahead.
In the same set of predictions, one of those codenames was said to be "Opus 5.1". There is no Opus 5.1. There was never an Opus 5.1. More than three weeks later the docs still go straight from Opus 5 to nothing, and the model that actually shipped alongside Fable 5.1 was Mythos 5.1.
That is the pattern to carry into this week. Codename leaks that reach a real artifact have been a decent timing signal for Anthropic launches. The mapping from codename to product name has been wrong at least as often as it has been right, because the people reading the codenames have no way to know which family a stealth-tested checkpoint belongs to. So "a model is coming soon" is a reasonable read. "It is called Opus 5.5 and costs $4" is not.
Was the outage a launch signal?
It was not. Anthropic's status page records an incident of elevated errors across multiple models running from 00:50 to 02:10 UTC on September 22, 2026, affecting Fable 5, Fable 5.1, Mythos 5, Mythos 5.1 and Opus 5. The affected list names the models that already exist.
Reading a serving incident as deployment activity is a durable piece of AI-community folklore and it almost never holds up. Capacity gets reshuffled constantly, and a launch day is in fact one of the times a provider most wants its existing endpoints to look calm. The honest read is that an incident happened on the same night as a rumor and the two got stapled together.
What Anthropic's release cadence actually suggests
Strip out the leaks and there is one signal left that has predictive value: how often Anthropic ships. Through 2026 the gaps between Opus releases have run roughly six to ten weeks, with Opus 4.7 to 4.8 near six weeks and 4.8 to Opus 5 near eight. Opus 5 landed on July 24, 2026. Fable 5.1 and Mythos 5.1 landed on September 1.
Two months on from Opus 5, an Opus-tier refresh in late September or October is entirely consistent with that rhythm. That is the strongest defensible statement available today, and it is a base rate rather than a schedule. It also has nothing to say about the version number, the price or the benchmark position, which is the part everyone actually wants.
For context on what the current generation already does, our Opus 4.8 review and Fable 5 review track how the tiers have moved release to release, and what Opus 5 did inside a real security research chain in July is a better measure of the jump than any leaked table.
Will it even be called Opus 5.5?
Probably worth assuming not, or at least not assuming so. The live sequence has been Opus 4.5, 4.6, 4.7, 4.8, then 5. A half-step to 5.5 straight off a major version is not a pattern Anthropic has used, while Fable went 5 to 5.1 and the rumor mill's own last attempt at naming produced a non-existent Opus 5.1.
If you are writing internal documentation, pinning a model ID in a config, or planning a migration around this, plan around a capability arriving rather than a string. Which is the practical point of this whole piece.
How to read an AI model rumor in five minutes
This cycle will repeat in six weeks with a different codename, so here is the check we actually run before covering one.
- Look for the string in a first-party surface. Open the vendor's model list, the API release notes and the pricing page. A model that is in limited release still usually exists as a row somewhere.
- Then look in third-party surfaces. Routers, IDE model pickers and leaderboards get new model IDs early and leak them by accident. This is where July's Honeycomb sighting came from.
- Check the numbers against the current lineup. A leaked spec that exactly matches what already ships, or that is a round percentage off a published price, deserves suspicion rather than a headline.
- Count the independent sources. Twenty accounts repeating one anonymous screenshot is one source. Ask who saw the thing rather than who posted about it.
- Treat market odds as sentiment. Read the resolution criteria, notice what they exclude, and never cite a market price as evidence of the event it is pricing.
- Fall back to cadence. Release rhythm is boring, public and has beaten leaks for accuracy all year.
What to expect, and what this changes for your stack
The reasonable expectation: a new Anthropic frontier model within weeks rather than months, arriving with an announcement, a model card, docs and a price, the way every previous one has. It may be called Opus 5.5. It may be 5.1, or 5.2, or something in the Fable line instead. The circulating price list may turn out to be close, and if it does, that will be a coincidence dressed as a scoop.
What it changes operationally is less than the discourse implies, and this is where we would push back on the excitement. If a 20% price cut on your frontier tier is material to your unit economics, the model ID is not your problem, your architecture is. Teams that route through an abstraction, keep evals in version control and can swap a model behind a flag got Opus 5 into production in an afternoon in July. Teams that hard-coded a model string spent that week on a migration instead.
Build so the next model launch is a config change
Every six weeks there is a new frontier model, a new price list and a new set of claims about it. The teams that benefit are the ones who can swap the model, re-run their evals and see the answer the same day. The teams that suffer are the ones who discover their prompts were tuned to one checkpoint's quirks.
Valletta Software Development builds the layer that makes this boring: a provider-agnostic routing layer, an eval suite that runs in CI against your own tasks and data, cost and latency telemetry per route, and a migration path that does not require touching product code. Tell us which models you are on and what you are spending, and we will tell you what a swap would cost you and what it would save.
FAQ
Is Claude Opus 5.5 released?
No. As of September 22, 2026 there is no Claude Opus 5.5. Anthropic has not announced it, its model overview lists Fable 5.1, Mythos 5.1, Opus 5, Sonnet 5 and Haiku 4.5 as the current lineup, and the September API release notes contain no new Opus model. Every claim about it currently traces back to posts on X.
What is the Claude Opus 5.5 release date?
There is no confirmed release date. Posts on X pointed at September 22 or 23, 2026, and one of the most-quoted accounts moved its own prediction by a day within hours. The defensible statement is the cadence one: Anthropic's Opus releases have been about six to ten weeks apart in 2026 and Opus 5 shipped on July 24, so a late September or October refresh is plausible without being scheduled.
How much will Claude Opus 5.5 cost?
Unknown. The circulating figure is $4 per million input tokens and $20 per million output, with a $0.20 cache read, which is a clean 20% below the documented Opus 5 pricing of $5 and $25 with cache reads at 10% of input. No source has been attached to that table, so treat it as a plausible guess rather than a leak.
What is claude-wafer-eap?
It is a codename asserted in posts on X to be Anthropic's internal name for a stealth-tested Opus 5.5. We could not find the string in any verifiable artifact: not in Anthropic's docs or API, not in OpenRouter's model listings, not on a public leaderboard. The -eap suffix does match the pattern of codenames seen before previous Anthropic launches, which is why the claim sounds plausible.
Will Claude Opus 5.5 have a 1M token context window?
The claim of a 900K to 1M context window would not be an upgrade. Opus 5, Sonnet 5, Fable 5.1 and Mythos 5.1 all already document 1M token context windows, with 128K maximum output. A leak whose headline spec matches what already ships is weak evidence of a real spec sheet.
Was the September 22 Claude outage related to a launch?
No. Anthropic's status page logs elevated errors across multiple models from 00:50 to 02:10 UTC on September 22, 2026, affecting Fable 5, Fable 5.1, Mythos 5, Mythos 5.1 and Opus 5. Those are the models that already exist, and reading a serving incident as deployment activity has a poor track record.
Will it be called Opus 5.5, 5.2 or 5.1?
Nobody outside Anthropic knows, and the rumor mill's naming record is bad. Late August predictions correctly called the September 1 launch date one day ahead while claiming one of the two models would be Opus 5.1. No Opus 5.1 has ever shipped; the second model was Mythos 5.1. Plan for a capability arriving, not for a specific model string.
Is Opus the most capable Claude model?
Not currently. Fable 5.1 and Mythos 5.1 sit above Opus 5 in Anthropic's range at $10 input and $50 output per million tokens, double Opus 5's $5 and $25. Part of what makes the Opus 5.5 rumor eye-catching is that it claims the next Opus will match that premium tier at a fraction of the price, which would mean Anthropic collapsing its own price ladder.