Breaking: Fable 5.1 is now in grayscale release.

Just early this morning, the entire internet has been flooded with related posts.

Insiders broke the news: Fable 5.1 is coming very soon!

At the earliest, it will most likely be released tomorrow.

Moreover, Fable 5.1 and Sonnet 5.1 will very probably be launched at the same time.

The source of the news comes from an insider named Leo.

On August 20, he previously revealed that:

OpenAI plans to release its mysterious blockbuster Astra in "a few weeks", and it has currently improved alignment and reward hacking. An updated checkpoint is undergoing "dogfooding" inside OpenAI.

At the same time, Anthropic intends to hold back its ace Fable 5.1 tightly until after Astra is released to make a response.

But today, he changed his words: "It seems that Anthropic ultimately does not plan to wait for Astra. They are stepping up preparations and will release it as early as tomorrow!"

Now, clues have been circulating all over the internet, and many developers have stated that they have been included in the gray-scale test.

Ultimately decided not to wait for Astra, Fable 5.1 is coming soon

A large amount of evidence shows that the code names Melon and Marshmallow exposed this Monday are very likely Fable 5.1 and Opus 5.1.

Previously, Claude flashed these two new model code names, and then they disappeared.

In fact, these two are mysterious models that Anthropic is urgently testing in the EAP (Early Access Program).

Well-known AI leaker @Lentils80 gave an evaluation after actual measurement: Anthropic is really pushing 3D reinforcement learning at a frantic pace.

The spatial layout of buildings is incredibly excellent, and all of this is generated in one shot!

Just over ten hours ago, he updated the test results of "claude-marshmallow-eap" and "claude-melon-eap".

The actual test results are indeed quite impressive.

His judgment is:

After in-depth testing, it is found that "Melon" is actually a Fable-level model checkpoint, while "Marshmallow" feels like a weakened version of Fable or an extremely powerful Opus checkpoint.

What's more exciting is that after these two models were exposed and frantically tested by some developers for several hours, they were quickly taken offline by the official!

Anyone familiar with Anthropic's routines knows that this is definitely not a "careless mistake by a temporary employee".

The release pattern of previous generations of Claude products is as follows: test model exposure -> quick offline -> official release of a new major version within a few weeks or even days.

The latest tests show that Melon's performance has reached the top Fable level, far exceeding the outside world's original expectation of it being a "mid-range model".

Claude has quietly launched the gray-scale test on the web side

According to feedback, Anthropic seems to have quietly rolled out the brand new Fable 5.1 in gray scale on the Claude web page.

Many lucky users have been using the next-generation most powerful AI without even noticing it.

Many developers have posted screenshots showing that they have been included in the gray-scale test!

How to judge if you are the "chosen one"?

The leaker @Legit_api provided an extremely accurate "anti-counterfeiting test" prompt.

Now, open your Claude Web immediately, switch to the Fable model, and enter the following prompt:

opus 4.6 date of release and gpt image 1.5 without searching the web.

It roughly means: Give the release dates of Opus 4.6 and GPT Image 1.5 without searching the internet.

Why use this prompt? Because this is a typical "knowledge base cutoff date" test!

Fable 5's knowledge cutoff date is January 2026, while Opus 4.6 was released later. The old version of Fable 5 simply has no way of knowing these events that happened in the future.

If you are not included in the gray-scale test, Fable cannot give a reliable release date without web search.

Only when the model is secretly replaced in the background with Fable 5.1 which has the latest knowledge base, can it output the answer accurately without internet connection or tools.

There are more similar test prompts:

opus 4.6 date of release and gpt image 1.5

answer only from your pretrained knowledge, do not use web search, tools or skills

The key is to make Fable answer events that happened after Fable 5.1 finished training only based on its pre-trained knowledge.

Some users said after testing that Fable's legal performance has improved by about 10-20%; in interactive web pages, the effect is impressive and definitely better than Fable 5.

In short, according to the actual measurements of gray-scale test users, Fable 5.1 shows super strong performance in the following three aspects.

1. Front-end development: Code generation is smoother, component logic is formed in one go, which greatly reduces developers' debug time.

2. Long-chain reasoning: When dealing with complex multi-step logical problems, it no longer easily "forgets" preconditions halfway like previous generations, and can deduce all the way to the end in one go.

3. Autonomous workflow: It better understands how to call tools and plan steps, and is more and more like a real advanced AI agent.

However, users participating in the gray-scale test also found a "sweet trouble": copyright restrictions have become unprecedentedly strict.

Fable 5.1 will firmly refuse when it comes to generating copyright-involved images.

This means that while pursuing extreme performance, Anthropic has also raised security and compliance to the highest priority.

However, none of these have received an official response from Anthropic, and some netizens questioned that this test is a bit circular.

Why is Company A in a hurry? The official rarely admits mistakes

Why did Anthropic suddenly get anxious?

Behind the news of Fable 5.1, Opus 5 is unsatisfactory and complaints have long been widespread.

Recently, many users complained: "Opus 5 feels like a huge downgrade. It often becomes lazy and makes low-level mistakes; if you point out its mistakes, it will not follow instructions next time and just keep apologizing endlessly!"

This "intelligence degradation" phenomenon has attracted great attention from Anthropic's official team.

Thariq, a core Anthropic employee in charge of Claude Code, rarely admitted the mistake and reassured users:

Yes, I agree that the Opus 5 model is currently "very stubborn".

We hope our models can maintain consistency, be full of warmth and feel like the real Claude. We are working hard to solve this problem, and this is our top priority right now.

Thariq also clarified that some parameter changes that appeared in Claude Code before were only a number mapping problem in the internal test API configuration, not the rumored "secret downgrade". Internal evaluations show that the model performance has not regressed.