Claude Fable 5.1: Leak, Date, Price, What Holds Up

An alleged leak about Claude Fable 5.1 has been circulating for a few days: launch in August, pricing unchanged. I traced the sources back, and the result is more useful than the report itself, because it produces a test you can reuse the next time a rumor wave starts. The video version of this analysis sits at the top of this page and on our video page.

First the separation that most reports skip. I keep three things strictly apart here: what I checked myself in Anthropic's official documentation, what Anthropic states as the vendor about its own models, and what posts and outlets claim. Which models we actually use in the agency we keep transparent on our AI in numbers page.

The negative finding: no entry, no model ID

First the part that can be checked. On 29 July 2026 I opened Anthropic's official model overview and the platform release notes. On Claude Fable 5.1 there is nothing: no model name, no model ID, no entry. Four current models are listed, plus the legacy line.

One detail makes that finding sturdier than a plain keyword search. The model overview lists, for every model, the identifiers for Amazon Bedrock and Google Cloud. Anyone putting a model into production needs those rows, and they sit publicly in the same table. For a Fable 5.1 there is none.

That is not proof that nothing is coming. It is proof that nothing was announced at the time of the check. The whole article rests on that difference.

Strand one: the Mythos post from 21 June

The story has two strands, and they are not equally strong.

On 21 June 2026 Andrew Curran wrote on X, in substance, that training had produced a more capable version of Mythos, possibly as Mythos 5.1, possibly as Mythos 6.

Why that matters: according to Anthropic's model overview, Claude Mythos 5 sits in the same class as Fable 5, with identical specifications and identical pricing, only it is not openly available but invitation only within Project Glasswing. A stronger Mythos would therefore genuinely point to a stronger model of that class.

Why it is still not evidence: the post names Mythos, not Fable 5.1. A "maybe Mythos 5.1, maybe Mythos 6" does not turn into a product name, let alone a date.

In late July Curran followed up and wrote that he thinks 5.1 is finished and that Anthropic is holding it back for the next OpenAI release. The post frames that explicitly as his own guess, and that is exactly how I pass it on.

Strand two: a single account, two posts

The second strand comes from a single X account. It mentioned a Fable successor for August as early as 14 July 2026. On 25 July the claim gained a name and a price: Fable 5.1, targeting August, pricing unchanged. Where the information came from, the post does not say.

I deliberately pass no judgment on the person behind the account. A date circulating without a named origin is exactly that: a claim without a verifiable basis. That is a statement about the evidence, not about intent.

The echo: how one claim becomes many reports

Then comes the thing that makes every rumor wave big: the echo. Several sites reproduce the same content, often with phrasings such as "multiple reports suggest". The outlet thewincentral words it precisely that way and notes in the same text that Anthropic has not officially confirmed these details yet.

One example shows the mechanism especially clearly. The article at 36kr claims a launch in its headline, while its own body text says the model is scheduled for release in August. Headline and text on the same page do not agree.

On 28 July geeky-gadgets additionally added beta testing through a red team portal. The source the article gives for that is a YouTube video. For a source check that means: the number of reports grows, the evidence does not grow with it.

The counter voices

They exist, and they get quoted less often. Evolink states verbatim: "Anthropic has not announced a model named Claude Fable 5.1", and on the circulated pricing: "Current Fable 5 pricing cannot be carried forward as a Fable 5.1 fact."

AIToolsReview goes a step further and does consider a hardened point release in the Mythos class a conceivable next step, but writes in the same breath that nothing has been announced. That is the clean position: plausible and unconfirmed at the same time.

The documented counter case: Claude Honeycomb EAP

So that none of this reads as blanket leak skepticism, here is the counter case from the same month, one that actually worked.

On 8 July 2026 an entry called "Claude Honeycomb EAP" stood in the model picker of the Cursor development environment. According to The New Stack it came with descriptive text, an early access label, a one-million-token context window and a safety fallback to Claude Opus 4.8. The entry disappeared again within hours, and Anthropic never commented on it.

Two weeks later, on 24 July 2026, Anthropic released Claude Opus 5 with the model ID claude-opus-5. The details from the Cursor entry matched it, so the sources report. I attribute that match explicitly to those reports and not to a check of my own, because a one-million-token context window and a high effort level exist on other models too. What Opus 5 actually delivers in benchmarks and price I broke down separately in the post on the Claude Opus 5 benchmarks.

The test: real leaks leave artifacts

What made the Honeycomb case a leak was not the spec list but an unreleased codename in a production interface. Honesty requires adding this: the sighting traces back to a single developer who posted the screenshots on X on 9 July, and every newsroom is working from the same images.

That makes the point stronger, not weaker. The difference between Honeycomb and Fable 5.1 is not the number of sources. It is the existence of a verifiable artifact.

Out of that comes a rule you can apply yourself to the next report. Real leaks leave artifacts: a string, a screenshot, concrete details in an interface, a model ID in documentation. Claims without an artifact remain claims, even when ten sites repeat them. The same caution already paid off in the benchmark and cheating debate around GPT-5.6 Sol.

Staying fair: the counter check in both directions

Anyone checking strictly has to check in the other direction as well. The same account that named the August date announced a Thursday launch one day before the Opus 5 release. Opus 5 arrived on Friday the 24th of July. The prediction that the new model would not surpass Fable 5 was not confirmed either: on its own Opus 5 page, Anthropic reports a result on OSWorld 2.0 above Fable 5's best figure.

That does not automatically devalue any future statement. It only shows that a hit rate is no evidence here, in either direction.

The price question that actually counts today

The circulated 10 and 50 dollars per million input and output tokens are not leaked. Those are today's prices for Claude Fable 5, simply projected forward in the reports. Evolink says the same.

The number that matters in practice is a different one, and it is documented: according to Anthropic's pricing figures, Claude Opus 5 costs 5 dollars per million input tokens and 25 dollars per million output tokens, half of Fable 5.

On capability, Anthropic itself says Opus 5 comes close to the frontier intelligence of Fable 5, at half the price, and cites for OSWorld 2.0 a result above Fable 5's best figure. CNBC reports exactly these vendor numbers. None of that has been independently verified, and I therefore label it explicitly as vendor-reported. How far vendor figures and independent measurements can drift apart is also visible in the story around the comeback of Fable 5 and in the comparison with the open model Kimi K3.

My take

Now the assessment, and this is my opinion based on the points documented above, not an external fact.

A point release above Fable 5 strikes me as plausible. According to the Anthropic release timeline by hidekazu-konishi, the gaps between this year's Opus point releases were 73, 70, 42 and 57 days, so six to a good ten weeks. I do not read a clear trend into that, and a release date certainly does not follow from it. Plausible simply is not announced.

For practice that means two things from my experience. First: do not plan a migration onto a model that has no model ID. For your operation, a model only exists once it appears in the documentation and in your cloud's catalog. Second: the more interesting lever right now is not the next version but the models already available, which most companies are nowhere near exhausting. If you want to work out which model fits your tasks, your data protection framework and your budget, that is exactly the subject of our AI consulting.

I will report back as soon as an artifact turns up. Until then the finding of 29 July 2026 stands: no announcement, no model ID, no entry.

Sources

• Anthropic, model overview: platform.claude.com/docs/en/about-claude/models/overview
• Anthropic, platform release notes: platform.claude.com/docs/en/release-notes/api
• Anthropic, Claude Opus 5 announcement: anthropic.com/news/claude-opus-5
• The New Stack on Claude Honeycomb EAP: thenewstack.io
• Evolink on the announcement status and the pricing: evolink.ai
• AIToolsReview on the next Claude model: aitoolsreview.co.uk
• 36kr, headline versus its own body text: eu.36kr.com
• thewincentral on the "multiple reports" phrasing: thewincentral.com
• geeky-gadgets on the beta testing addition of 28 July 2026: geeky-gadgets.com
• CNBC on the Opus 5 vendor figures: cnbc.com
• Andrew Curran on X, 21 June 2026: x.com and late July 2026: x.com
• The three posts from the single account: 14 July 2026, 23 July 2026, 25 July 2026
• Anthropic release timeline: hidekazu-konishi.com

As of: July 2026

Frequently asked questions

Has Claude Fable 5.1 been officially announced?

No. As of 29 July 2026, neither Anthropic's official model overview nor the platform release notes carry an entry for Claude Fable 5.1: no model name, no model ID, no row among the identifiers for Amazon Bedrock and Google Cloud. Four current models are listed plus the legacy line. The outlet Evolink also states explicitly that Anthropic has not announced a model by that name.

Where does the August date for Fable 5.1 come from?

From two strands. On 21 June 2026 Andrew Curran wrote on X, in substance, that training had produced a more capable version of Mythos, possibly as Mythos 5.1, possibly as Mythos 6. Separately, a single X account mentioned a Fable successor for August as early as 14 July and on 25 July added the name Fable 5.1 together with unchanged pricing. Where that information came from, the post does not say.

Are the quoted 10 and 50 dollars leaked prices?

No. Those are today's prices for Claude Fable 5 per million input and output tokens, simply carried forward in the reports. Evolink puts it this way: current Fable 5 pricing cannot be carried forward as a Fable 5.1 fact. What is documented, and practically more relevant, is the price of Claude Opus 5: 5 dollars for input and 25 dollars for output per million tokens, half of Fable 5.

What was Claude Honeycomb EAP?

An entry that, according to The New Stack, briefly appeared in the model picker of the Cursor development environment on 8 July 2026, with descriptive text, an early access label, a one-million-token context window and a safety fallback to Claude Opus 4.8. The entry vanished within hours, Anthropic never commented on it, and the sighting traces back to a single developer who posted the screenshots on X on 9 July. Two weeks later Claude Opus 5 appeared, and the details matched it according to the reports.

How do I recognize a real AI leak?

By the verifiable artifact. A real leak leaves something you can follow up: an unreleased codename in a production interface, a screenshot, concrete details in a user interface, a model ID in documentation. What does not decide it is the number of sites reporting, because reports copy one another. Where no artifact exists and the official documentation carries no row, a claim stays a claim.