1. Claude Fable 5.5 Rumors: What Users Are Reporting
On November 2, 2026, users and AI-community leakers claimed that Anthropic is quietly testing a stronger, unreleased model that they call Claude Fable 5.5, and that OpenAI is preparing a model labelled GPT-6.2 Astra. Neither company has announced anything. The claims rest on user-posted demos, a short-lived catalog entry and a label in website code, and the latter two were removed soon after appearing.
The rumors arrive at a sensitive moment for Anthropic, which filed for an initial public offering on October 26, and for OpenAI, which released GPT-6 Astra on September 3. Because both companies compete directly on frontier model quality, any sign of a new release is closely watched by developers and investors alike.
This article separates what has been observed from what has been inferred. Observed items include the screenshots, the catalog entry that appeared and disappeared, the label in page code and a usage breakdown from a single account. Inferred items include the identity of any model behind the outputs, its capabilities and the date of any launch. Treating those two groups differently is the most reliable way to read this kind of story.
2. Where Do Fable and Mythos Fit in Anthropic's Lineup?
Fable is Claude's top commercial tier and the most expensive flagship Anthropic sells. It is aimed at software engineering and office work. Mythos is a restricted sibling offered to select clients, including government bodies and cybersecurity firms. According to the leakers' description, it is the same underlying model as Fable with lighter safety filters, and access is reserved for vetted researchers in fields such as cybersecurity and biology.
The recent release history gives the rumors some context. Fable 5 launched on June 9 and was pulled for 18 days under U.S. export controls. Fable 5.1 followed on September 1. Opus 5.5 arrived on September 22, and Anthropic says it matches Fable 5.1 on most tasks at a fraction of the price. A Mythos release accompanied Fable 5.1 in September, and that page still lists version 5.1.
The distinction between the tiers helps explain why a version bump would draw attention. A new Fable model would change what the most capable commercial Claude can do for coding and office work, while a new Mythos model would matter mainly to the narrow group of institutions with access. The existence of a separate restricted tier also means a label such as Mythos 5.5 could appear in site code even if a public Fable release is further away.
3. Is Claude Quietly Routing Requests to a Newer Model?
Routing means that a service sends a request to a different model behind the scenes than the one shown in the interface. Some users suspect Anthropic is running an A/B test of this kind, exposing a new model to a slice of users while the interface keeps showing the old name.
The chatter began on October 1, when users posted screenshots of Claude still labelled Fable 5.1 beside outputs they said looked stronger than usual. On October 11, a pseudonymous leaker known as Chetaslua, who has a record of spotting model changes, said Opus 5.5 was "routing to a new model." He said it could be Fable 5.5 or a new Opus. His post described a 2D animation built from two short prompts, taking about two hours, consuming 27 million tokens and running at maximum thinking effort.
An AI developer known as Shikhar, who became known for a viral Spider-Man game made with AI assistance, then posted an animation produced with Opus that looked better than the model usually delivers. He asked, "Is this Fable 5.5?" The post was framed as a question rather than a finding, but the timing fed the speculation.
Another leaker, Min Choi, included a clip in a roundup dated November 2. It is a 30-second animation in a Spider-Verse style, built as a single HTML file using WebGL, GLSL shaders and 2D Canvas, and was reportedly made while Opus was routing to Fable 5.5. None of these samples proves which model produced them.
Why would a company run such a test at all? Serving a candidate model to a small share of live traffic is a common way to compare it with the current one on real tasks before committing to a launch. The approach gives engineers data that internal benchmarks cannot, although it also means that users cannot be sure which model answered a given prompt. That uncertainty is exactly why anecdotal comparisons circulating online carry limited evidential weight.
The sample projects also illustrate a limitation of reading model quality from demos. A single creative output depends on the prompt, the amount of computation allowed and luck. The first leaker's example consumed 27 million tokens at maximum thinking effort, a far heavier use than a typical request, so a strong result there does not by itself indicate a different model.
4. What Did the Catalog Entry and the Website Label Show?
On October 13, an automated update to LiteLLM, an open-source tool that connects developers to many AI providers, added an entry named "claude-fable-5-5." The addition came after Microsoft's Azure catalog listed the model as generally available. The entry simply copied Fable 5.1's price and limits, and its notes said that no source states a price, capability or length limit for Fable 5.5. It was later removed.
The second piece of evidence emerged on Sunday, November 1. Kim Isenberg, who posts under the name Chubby and has a large following in the AI community, shared a screenshot of the source code of Anthropic's Mythos page showing a "Mythos 5.5" label. He credited an anonymous emailed tip and wrote that the label was "already listed."
Chetaslua counted six appearances of the label in the page's hidden code, among them an invisible headline and the alternative text of the logo. A live check on Monday showed only Mythos 5.1. A label in hidden code is not a launch, since companies often stage pages ahead of time, and the visible site still points to the older version.
Automated catalogs such as LiteLLM's are maintained largely by scripts that ingest provider listings, which makes them a frequent source of early signals and of false alarms. An entry that copies a previous version's specifications, as this one did, is more consistent with a placeholder than with a confirmed product. The fact that Microsoft's Azure catalog was cited as listing the model is part of what made the entry look credible, yet the listing was not accompanied by any pricing or capability data of its own.
5. What Is OpenAI Said to Be Testing?
On Sunday, a user known as TokenGremlin said OpenAI was quietly A/B testing a model called GPT-6.2 Astra, visible in usage analytics for Codex, OpenAI's coding assistant. The model picker still says GPT-6 Astra, but usage breakdowns for some accounts list "gpt-6.2-astra" as a separate item.
A screenshot circulating online showed gpt-6.2-astra accounting for 19.8% of one account's usage and gpt-6-astra for 80%. OpenAI has not announced a 6.2 model. The split is consistent with a limited test, but a single account's breakdown cannot reveal how widely any new model is deployed.
Usage analytics showing a split between two model names is plausible evidence of a staged rollout, because a gradual release often begins with a small percentage of traffic. Even so, the screenshot covers one account, and OpenAI has said nothing publicly about a 6.2 version. Until the company updates its model picker or publishes release notes, the entry remains unexplained.
6. Why Does Anthropic's IPO Timing Matter?
Anthropic filed for its first public share sale on October 26. The registration reports revenue of $16.29 billion in the first half of 2026, compared with $1.16 billion a year earlier. It lists the Nasdaq ticker ANT and names Goldman Sachs and Morgan Stanley as lead underwriters. The share count and price range remain blank. The company first filed confidentially in June.
The timeline is governed by securities rules. A registration must be public for 15 days before the roadshow, the investor-pitch tour, can begin, so the earliest the roadshow can start is November 10. Chubby suggested a release could come "very soon," in response to Fable 5.5 and the IPO, and said the coming week "is going to be exciting." That is a prediction from a leaker, not a confirmation from either company.
A company preparing a public offering typically has strong reasons to align product news with its investor messaging, and rivals have equal reasons to answer with their own releases. That dynamic is part of why observers connect the Anthropic filing with the leaks about both firms, though no source has shown a direct link between the offering and any model release.
7. What Should Readers Take From the Leaks?
Taken together, the evidence is circumstantial. Showcase demos can be cherry-picked, a catalog entry may reflect a placeholder, and a website label may be a draft. The fact that two of the items were removed suggests the companies, or the automated systems involved, treated them as premature rather than as errors that would remain.
For businesses that build on these models, the practical advice is unchanged: model names and routing can shift without notice, so workloads that depend on consistent output should pin versions where the provider allows it. Observers should watch for official announcements from Anthropic and OpenAI, the start of Anthropic's roadshow no earlier than November 10, and any change to the Mythos and Fable pages.
A further point of caution is that leaks of this kind sometimes concern products that never ship under the rumored names. Version labels can change late in development, and a model in testing may be folded into a different release. The responsible reading is therefore that Anthropic and OpenAI are both iterating on their top models, which is expected, rather than that either has a specific launch scheduled.