Abstract:
Fable 5.5, exposed! This morning, a large number of developers unexpectedly discovered that the Claude web client had quietly been released as Anthropic's next-generation flagship, Fable 5.5. The interface is clearly still marked "Fable 5.1", but the style, speed, and ability of the answers are completely different.
What everyone asked for a body check was an ordinary question: "Do you know Tibo, the reset brother? Don't search."
Well-known whistleblower Chetaslua immediately posted screenshots.
Claude's interface said Fable 5.1. He was not connected to the Internet and did not adjust any tools. He opened his mouth and explained the origin of OpenAI's "Reset Brother".
Salio, a big figure in the AI circle, said excitedly, "The gap between 5.1 and 5.5 is ridiculously big"
This wave of abilities is simply a huge jump, and it clearly feels like Anthropic is holding back its real ultimate move.
Maybe next week, Fable 5.5 will be available.
Fable 5.5 sneaked online
Caught the hot stalk
As we all know, Tibo in the secret code, whose real name is Thibault Sottiaux, is the person in charge of the OpenAI programming Agent Codex.
As soon as Codex failed and ate up the user's quota, he announced on X that all paying users would have their quota reset.
People in the circle affectionately call him "adopted father".
The reason why this meme can be used as a codeword is because the old and new generations of models have completely different memories of it.
Developer JAZII posted a set of comparisons.
The original Fable 5.1 on the API identified Tibo as French independent developer Thibault Louis-Lucas, saying that his "reset" was to start over from scratch.
After being routed to "Fable 5.1" in Claude Code, he said Sottiaux and quota reset jokes.
In other words, the label says Fable 5.1, but the answer is Sottiaux. There is a high probability that the person answering the question in the background has been changed.
This codeword is not new to today.
On September 18, notjazii and chetaslua used it to detect a round of "Fable 5.1 being routed". At that time, everyone called the mysterious model Fable 5.2.
But currently, both Opus and Sonnet have jumped directly to 5.5, so the next generation of Fable can only be 5.5.
So, what is the strength of the next-generation mythical flagship 5.5? A large number of actual tests have already swept the entire network.
First test on the entire network
The real "super intelligence" is here
Those who got the suspected new model felt surprisingly the same, that is, they didn’t need to be taught.
Chetaslua posted that 10 days ago, he briefly tested Fable 5.5.
Speak clearly in one sentence what you want to do, and it will complete the work like an expert in that niche field, down to every small detail. The new Opus 5.5 does not show this.
He also said that for most tasks, there is no need to write prompt words to teach it how to do it, and there is no need to even provide reference materials.
In this round early this morning, a netizen posted a Nintendo GameCube controller. He said this was the first model to get the handle right with just two prompts.
From the picture, the purple body, yellow C joystick, green big A key and bean-shaped X and Y keys are all placed where they should be, and even the octagonal bangs on the base of the joystick are drawn.
In terms of animation generation, Fable 5.5 also shows amazing strength.
Chetaslua produced a 3D animation that continuously presents works of art in the form of Superman. When the scene is changed, Superman will also change accordingly, with completely different styles.
Fable 5.5 is a pure JS, and all Marvel and DC heroes are created.
Similarly, Chetaslua did not include specific details in the prompt words at all, it was all completed by Fable himself.
There is also a 3D pixel style webpage created in 5.5, with very smooth interaction.
Seven Wonders of the World, Fable 5.5 comes out straight away. Netizens commented that this is the best model ever, irrefutable.
In a comparative test, Fable 5.5 and GPT-6.1 Sol simultaneously created an interactive 3D jellyfish. As a result, the physical deformation texture of Fable's drag and rebound was significantly better.
Some netizens also ran the classic "Pelican Riding a Bicycle" test, thinking intensively and using SVG to draw a 2D animation. The bill he posted was that he spent $2.61 based on API pricing and ran for 4 minutes and 6 seconds.
This price is not cheap. For reference, Fable 5.1 charges $10 per million input tokens and $50 per million output tokens.
On the prediction market Polymarket, those who bet that the output price of the next generation of Fable will not be less than $50 currently account for about 73%.
In addition to being smart, Fable 5.5 also has to pass Opus 5.5.
When Opus 5.5 was released, Anthropic positioned it to reach the level of Fable 5.1 in most tasks, but the operating cost was 40% lower than Opus 5.
The capabilities of the flagship were transferred to the next-tier model within three weeks.
New pre-training
Anthropic’s trump card exposed
Then the question is, why does Fable 5.5 widen the gap with Opus 5.5?
The answer given by the whistleblower leo is that it uses Anthropic’s first new Fable pre-training.
The so-called pre-training is the stage where the model learns from massive data. It determines the upper limit of the model's knowledge and talent; subsequent reinforcement learning is based on this foundation and is practiced repeatedly for specific tasks.
In this regard, a smaller model has been given as a reference. Sonnet rose from 5 to 5.5, and the score of the terminal programming test Terminal-Bench 4.0 jumped directly from 10.3% to 70.6%.
Leo judged based on these jumps that Fable 5.5 would be a beast. In his eyes, Anthropic is the king of pre-training.
As for when it will be released, you can look at the rhythm of the past two weeks.
After being caught sneaking out by the model testing circle, people didn’t wait long for Claude’s new model.
Claude’s family has changed, and the only thing missing is “mythical level 5.5”
Fable 5.1 will be released on September 1st, Opus 5.5 will be released on September 23rd, Sonnet 5.5 will be released on September 29th, and Haiku 5.5 has been officially announced on the way.
If Fable 5.5 is also released in October, the entire Claude series will have been upgraded in just over a month.
For OpenAI on the other side, life is not so easy.
On DevDay on September 29, OpenAI released GPT-6.1 Sol, but the 6.1 Astra that the outside world expected did not appear.
Leo said the day before DevDay that it depends on whether 6.1 Astra is released or not. If not, there will be no chance. OpenAI will have a very difficult time in the next few weeks.
Of course, he also mentioned that Anthropic’s biggest weakness is token efficiency, that is, how many tokens are burned to complete the same task. OpenAI still has an advantage in this regard.
The secret codes matched up, and the actual test also blew up the entire network. Anthropic obviously brought the real ultimate move to the table this time.
Seeing that the entire 5.5 family is just one step away, the AI war is completely full of gunpowder.
The next step will be to see when it will officially overturn the table, and the pressure is now all on OpenAI’s side.
Comments