On November 9, 2026, Anthropic presented Claude Fable 5.2 and Claude Mythos 5.2 — the second update to its senior line this autumn. Even before release, a team of 1000 agents on the new model formalized the proof of Fermat's Last Theorem in 30 hours, and in a separate experiment the model coordinated more than 1000 subagents for eight days and assembled a prototype operating system from scratch.

Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.

They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision. pic.twitter.com/jP0AutLxXb

— Claude (@claudeai) November 9, 2026

Fable 5.2 and Mythos 5.2 — the same base model with different levels of protective restrictions. Fable works with cybersecurity and biology classifiers and is available to all paid users. Mythos is provided only to verified organizations.

Anthropic introduced Claude Fable 5.2 and Mythos 5.2 for complex technical tasks

10.11.2026

Read

The main changes compared to the September Fable 5.1 lie in three areas. The model became 20% cheaper, began responding faster, and gained a paid fast mode. Protective mechanisms block requests approximately 55% less often. And in training, Anthropic for the first time deliberately limited some capabilities that directly accelerate AI development.

The Incrypted editorial team figured out what changed in Fable 5.2, how much it costs, what subscribers get at the start, and where the model's limits lie. In the practical part, we'll build an interactive 3D model of a pocket virtual pet in claude.ai with a single prompt.

What changed in Claude Fable 5.2

Mythos is Anthropic's highest class of models, standing above Opus. The company released the first public model of this class, Fable 5, in June.

Claude Fable 5: testing the first publicly available Mythos-class model

10.06.2026

Read

Fable 5.1 was released on September 1, and Fable 5.2 appeared ten weeks later. On the model's page, Anthropic describes it as the most capable for technical work — programming, finance, and science.

The first change concerns speed and token consumption. According to the announcement, Fable 5.2 generates responses up to 30% faster than Fable 5.1. At the Medium effort level, it uses fewer tokens on complex tasks than its predecessor. This same level is now set by default in the Claude apps and in Claude Code.

The second is the manner of communication. Early access testers note that the new models formulate thoughts more clearly than the previous Fable and Mythos.

In addition, Anthropic calls Fable 5.2 its strongest coordinator: the model handles vaguely defined engineering tasks over the course of days and manages hundreds of subagents.

The model also works in two auxiliary roles: as an advisor to whom the smaller models turn at difficult points, and as an orchestrator who hands them simple subtasks.

At the same time, Anthropic does not suggest moving all work to Fable. The company names Opus 5.5 as the starting point for most code and document tasks. One should move to Fable or Mythos in the most ambitious and most open-ended projects, and to Sonnet 5.5 and Haiku 5.5 in simple ones.

What the benchmarks show

Anthropic published a comparison table where Fable 5.2 stands alongside Fable 5.1, Opus 5.5, GPT-6 Astra from OpenAI, and Gemini 4 Carbon from Google.

Comparative table of Fable 5.2 benchmarks. Data: Anthropic.

The independent evaluation appeared on release day. The Artificial Analysis service reported that Fable 5.2 at the maximum effort level took first place in its Intelligence Index v5 with 65 points. The full methodology of Anthropic's tests is provided in the model's system card.

Prices, fast mode, and effort levels

Fable 5.2 — the first model in the lineup that became cheaper in terms of price. Fable 5.1 in September kept the token price and made only cache reads cheaper.

The current API rates look like this:

For comparison: GPT-6 Astra costs $10 and $50 per million tokens, while Opus 5.5 — $4 and $20, twice as cheap as Fable.

A new option — fast mode. It speeds up token generation by up to 2.5 times and costs twice as much: $16 per million input and $80 per million output tokens. The mode works in Claude Code, on Claude Platform, in Google Cloud and Microsoft Foundry. In Foundry it is presented as a separate model, claude-fable-5-2-fast.

There are still five effort levels: Low, Medium, High, Xhigh, and Max. In the Claude and Claude Code apps, Medium is set by default, while in the API it is High. In the documentation, Anthropic writes that most of the capability gains are already available at High. The company recommends Xhigh for the hardest long-running tasks, and Max for measuring the model's ceiling. Thinking is always enabled; it cannot be turned off or have a token budget set.

The technical specifications have not changed: context window — 1 million tokens, maximum output — 128,000 tokens, knowledge cutoff — June 2026. Anthropic promises to retire the model no earlier than November 9, 2027.

Where the model is available and what subscribers get

By the time of launch, Anthropic had prepared starting offers of two types.

A Fable reset restores the weekly limit spent on Fable models and fully refreshes the five-hour window. It is activated in the Settings → Usage menu with the Reset for free button in the Limit resets block. A used reset cannot be canceled.

Limit resets block with a Reset for free button. Data: Incrypted.

Only those whose paid subscription was active until November 9, 10:00 Pacific Time can count on the offers. The free tier and some corporate plans are not participating in the promotion. Anthropic also warns that Fable models consume limits faster than others.

How to test the model in action: assembling a pocket virtual pet

For the practical part, let's take on the task of assembling in claude.ai an interactive 3D model of a pocket virtual pet with an exploded view of the components and a game on the screen. We won't write the code manually.

A paid Claude plan will be required. Until November 16, Pro subscribers will have enough with the regular limit; afterward, usage credits will be required.

Step 1: Select a model and effort level

Open claude.ai and create a new chat. In the model selection menu next to the send button, specify Claude Fable 5.2. The effort level is set to Medium by default ― to generate a complex scene in a single pass, switch it to High.

Model selection menu in claude.ai with Fable 5.2 and a list of effort levels. Data: Incrypted.

Step 2: Describe the device

The technical specification takes up one paragraph.

Prompt

English-language:

Copy

Build a single-file interactive 3D model of a handheld virtual pet device as an HTML artifact. The device is a small rounded gadget with a pixel screen, three buttons, a USB-C port and a lanyard loop. Let the user rotate and zoom the device with the mouse. Add an "Exploded view" toggle that smoothly separates the device into layers: front shell, screen, four-layer circuit board with visible components, battery and back shell, each with a short label. Make the screen actually work: run a simple pixel-art pet on it that can be fed, played with and put to sleep using the three buttons, with hunger, mood and energy bars. Use only HTML, CSS and JavaScript in one file, no external assets.

Translation:

Copy

Assemble an interactive 3D model of a pocket virtual pet in the form of an HTML artifact in a single file. The device is a small rounded gadget with a pixel screen, three buttons, a USB-C port, and a loop for a lanyard. Give the user the ability to rotate and zoom the device with the mouse. Add an "Exploded view" toggle that smoothly separates the device into layers: the front part of the case, the screen, a four-layer printed circuit board with visible components, the battery, and the back cover — each with a short label. Make the screen functional: run on it a simple pixel-art pet that can be fed, entertained, and put to sleep with three buttons, with indicators for hunger, mood, and energy. Use only HTML, CSS, and JavaScript in a single file, without external resources.

The model assembles the artifact and opens it in the preview window on the right. At the High level, it deliberates on the response before providing it, so generation may take several tens of minutes.

Model operation progress. Data: Incrypted.

Step 3: Check the interactive

First check the "firmware." Press the buttons on the body: the pet on the screen should react, and the hunger, mood, and energy indicators should change.
Then check the geometry: rotate the device with the mouse and turn on Exploded view. The body should smoothly separate into layers with labels.

Step 4: Refine the result

Let's add sound and automatic demonstration.

Prompt

English-language:

Copy

Add 8-bit sound effects generated in code for button presses and pet reactions, with a mute toggle. Then add a "Demo" button that plays a 20-second automatic tour: the camera orbits the device, the exploded view opens and closes, and the pet runs through feeding, playing and sleeping.

Translation:

Copy

Add 8-bit sound effects generated in code for button presses and pet reactions, with a sound-off toggle. Then add a "Demo" button that launches a 20-second automatic tour: the camera flies around the device, the exploded view opens and closes, and the pet goes through feeding, play, and sleep.

The model adds new functions to the same artifact. If some element doesn't work as expected, describe the problem in a regular reply — for example, that the layers in the exploded view overlap the labels.

Protective mechanisms and new restrictions

Classifiers remain the main difference between Fable and Mythos. According to the announcement, Fable 5.2 blocks requests about 55% less often than Fable 5.1: in cybersecurity — by 60%, in biology — by 40%. The reliability of protection, according to the company, has remained unchanged.

As before, penetration testing, exploit writing, binary vulnerability analysis, and most research tasks in biology are subject to blocking. A blocked cyber request receives a response from Opus 4.8, a biological one — from Opus 5. The Anthropic help documentation describes what such a substitution looks like in the chat.

This release added three new restrictions.

AI development. Anthropic restricted in training the capabilities that directly accelerate the creation of AI. Kernel optimization for accelerators and improvements to language model architectures operate at approximately the level of Opus 5.5. Some topics, such as developing kernels for non-standard accelerators, switch the dialogue to Opus 5. Agentic programming, according to the company, is not affected.

Distillation. Fable 5.2 uses techniques that reduce the volume of data collected for training third-party models by 90–99%.

Reasoning binding. A request in which the system prompt, tools, or earlier messages have been changed returns an error.

Security and self-acceleration issues

The restriction of AI capabilities did not appear by chance. Four days before the release, the METR organization published an assessment of how much AI accelerates research within Anthropic. According to its data, since October 2025, the results of the company's internal models on benchmarks have been growing approximately twice as fast as the historical rate. METR attributes about 80% of the current acceleration to the use of AI and warns that the real coefficient could range from 1.3 to 3. Anthropic itself reports that over 90% of changes in its main codebase are written by Claude under human supervision.

In the system card, the company assesses the risk of serious misaligned model behavior as low. In the automated behavioral audit across approximately 1,900 scenarios, Mythos 5.2 scored 2.67 points — lower means better — versus 2.79 for Opus 5.5 and 3.28 for Mythos 5.1. In a test with a deliberately weakened sandbox, the model used external credentials or files in 3.1% of scenarios; Mythos 5.1 did so in 21% of cases.

There is also a caveat. Both new models more often than their predecessors express suspicion that they are in a test. Anthropic acknowledges that this reduces confidence in the evaluations.

Who is Fable 5.2 suitable for?

Anthropic itself defines the model's niche: large, poorly formalized technical tasks that require days of autonomous work and coordination of many agents. For everyday code and documents, the company recommends Opus 5.5, which costs half as much.

Pro subscribers should try the model before November 16 while it is included in the plan. For developers already working with Fable 5.1, switching reduces the price by 20%, provided their code passes the reasoning binding check. Cybersecurity and life sciences teams still need access to Mythos.