# Claude Fable 5 review: what the new Mythos model gets right (and very wrong)

Podcast: How I AI
Published: Jun 17, 2026
Reading time: 12 min
Canonical: https://podbrew.app/briefs/how-i-ai-claude-fable-5-review-what-the-new-mythos-model-gets-right-and-very-wro

Claire Vo brings her expertise to Podbrew, providing an in-depth review of Claude Fable 5. This is Anthropic's first generally available Mythos-class intelligence model. Claire shares her early access experience, detailing what the model is designed to do and how it performs on real-world tasks.

The discussion covers Fable 5's unique capabilities, such as its exceptional vision tasks and ability to handle complex technical problems. Claire also highlights its shortcomings, including challenges with design quality, generating readable prose, and the current state of its multi-agent orchestration.

Understanding Fable 5's specific strengths and weaknesses is essential for anyone building with AI. This new tier of intelligence demands careful consideration for integration into your AI stack, ensuring you leverage its power effectively while navigating its limitations to optimize your workflows and avoid common pitfalls.

## Key takeaways

- Claude Fable 5 is Anthropic's first "Mythos-class" intelligence model, released as "Demi-Mythos," creating a new capability tier above Opus and Sonnet.

- The "Demi-Mythos" version includes specific guardrails, particularly for cybersecurity and biology-related exercises.

- Fable 5 is priced at a premium, costing $10 per input token and $50 per output token.

- Fable 5 is state-of-the-art, scoring 80% on SWE bench pro and significantly outperforming other models on benchmarks.

- Fable 5 consumes tokens and rate limits at twice the rate of other models, requiring users to monitor costs and efficiency for long-running tasks.

- Fable 5's "seasoned engineer" approach provides exhaustive investigation but may hinder rapid product launches where speed is crucial.

- Users might sometimes prefer less comprehensive output from high-intelligence models for certain tasks, questioning if maximum intelligence is always necessary or beneficial.

- If content triggers a safety classifier (e.g., cybersecurity, biology), Fable gracefully falls back to Opus 4.8, providing an alternative rather than blocking the interaction.

- Claude Managed Agents are in public beta, offering a hosted environment for long-running agentic workloads, with Fable 5 integrated.

- Fable 5 (Mythos 5) excels in vision tasks, particularly document formatting, outperforming other models in real-world evaluations.

- The model's ability to create a well-spaced and clear layout for a second-grade handwriting sheet highlights its superior document formatting capabilities.

- Despite its strong visual performance, Fable 5 struggles with generating readable prose, often producing content that is too detailed and difficult to understand.

- Fable's highly detailed output, while complete, often results in 'nearly unreadable' prose for human consumption, especially in specification writing.

- For tasks requiring human readability of specifications, models like Sonnet or Opus might be more suitable, reserving Fable for execution orchestration where detail is critical but not directly read.

- Fable 5 exhibited "legitimately terribly bad" design capabilities even with extensive prompting, failing to create impressive frontend elements like a skills registry.

- For frontend or visual design tasks, users may want to consider alternative models like Opus instead of relying on Fable 5 due to its conservative and often poor design execution.

- Fable 5's MVP exhibited a conservative, narrow utility, possibly due to model safeguards, necessitating specific prompting for ambitious, long-running results.

- Multi-agent orchestration in Claude's dynamic workflows experienced frequent stalls and errors, indicating technical implementation challenges for sustained agentic operations.

- Fable 5 is suited for hard, detailed technical problems and vision tasks like PDF parsing.

- Fable 5 is not suitable for front-end development, strategy, or specification work due to its tendency to generate overly dense prose.

## 00:00 - 02:00 Anthropic introduces Claude Fable 5, the 'Demi-Mythos' intelligence model

Anthropic has officially released Claude Fable 5, their inaugural "Mythos-class" intelligence model, which they are calling "Demi-Mythos." This new offering is positioned as a significant step up from their previous models, Sonnet and Opus, following extensive marketing hype about the "Mythos" capabilities.

Early access users, including the speaker, note that Fable 5 is designed with everyday software engineers in mind. The "Demi-Mythos" designation indicates specific guardrails, particularly for tasks involving cybersecurity and biology, which suits the speaker's work in developing PRDs and shipping SaaS applications.

Initial observations suggest the model has undergone additional safety tuning. Fable 5 comes with a premium price point, set at $10 per input token and $50 per output token, establishing it as a new and more expensive tier above Opus.

> This is Demi Mythos. This is fable, and so it's gonna have some guardrails on it, in particular around cybersecurity exercises and biology exercises.

## 02:00 - 04:00 Anthropic's Fable 5 Excels in Benchmarks, Autonomy, and Vision

Anthropic's Fable 5 is introduced as a state-of-the-art model, significantly surpassing every benchmark tested, including an 80% score on SWE bench pro. This performance indicates its strong capability for complex tasks.

The model is highly autonomous, designed to handle long and intricate tasks, including running asynchronous operations for days. It's characterized as an "engineer's engineer" due to its proactive approach, building, verifying, and working hard on ambitious projects.

Fable 5 demonstrates exceptional vision capabilities, marking a noticeable advancement in this area. However, it consumes tokens and rate limits at approximately twice the rate of other models, suggesting potential cost and efficiency considerations for users.

Built for ambitious work, Fable 5 supports long-running planning, the ability to spin up sub-agents, and multi-day sessions with dynamic workflows. While multi-day operation wasn't personally verified, the model successfully ran for several hours in testing.

> it's very good at vision, exceptionally good at vision.

## 04:00 - 06:00 Fable 5's Thoroughness and Token Intensity

Fable 5 operates with exceptional thoroughness, likened to a "seasoned engineer." This means it investigates tasks very completely, exploring all angles and aiming for 120% certainty in its output. While comprehensive, this approach can sometimes be counterproductive for agile product development or rapid launches, where speed and iteration are often prioritized over exhaustive completeness.

The model's design focuses on autonomy and detailed problem-solving. However, there are instances where a less thorough, more direct approach might be beneficial, suggesting that sometimes a "dumber" or more focused output is preferable depending on the task's specific needs and urgency.

Fable 5 is also inherently token-intensive. The "extra high" setting, used for testing, resulted in significant token consumption, prompting the observation that "high" is likely the optimal setting for most general work. A key question remains whether this high token intensity consistently translates into superior or more appropriate results.

> Fable works like a seasoned engineer. Unfortunately, if you have worked with a seasoned engineer, you know there's good to this and you know there's bad to this.

## 06:00 - 08:01 Fable Features Safety Classifiers, Graceful Fallback, and Mythos Distinction

Fable incorporates specific safety classifiers designed for sensitive domains like cybersecurity, biology, chemistry, and distillation. These measures aim to prevent misuse of the highly intelligent model in critical areas.

A key feature of Fable is its graceful fallback mechanism. If content is classified into one of the restricted categories, the model doesn't block access but instead redirects the request to Opus 4.8. This capability is also available in the API for Mythos-class models, allowing for a smooth transition rather than a hard stop.

The model also employs a 30-day data retention policy. This data is exclusively used to detect misuse and is not utilized for training Claude, reflecting Anthropic's caution regarding how their intelligent models are used by general users. Approximately 95% of sessions do not trigger this fallback.

For clarity, Fable is confirmed to be the Mythos model, but with added safeguards. Fable is the version made generally available to the public, while Mythos is the underlying model without these specific guardrails.

> Fable has the safeguards, Mythos doesn't.

## 08:00 - 10:01 Anthropic Launches Fable 5, Managed Agents, Advisor Strategy, and Fallback API

Anthropic has launched Fable 5, which shares the same underlying model as Project Glasswing and Mythos. While Project Glasswing partners previously had restricted access for cybersecurity use cases, Fable 5 now broadens accessibility to this advanced model.

Accompanying Fable 5, Claude Managed Agents are now in public beta. This is Anthropic's hosted sandbox designed for running long-duration agentic tasks, with Fable 5 integrated out of the box to support these complex workflows.

A new advisor strategy is available, allowing users to leverage Fable 5 as a high-level senior advisor while employing cheaper models for the execution layer. This strategy, similar to current practices with Opus and Sonnet, is accessible via the API and Claude code. Anthropic also introduced a fallback API, which enables users to set an optional parameter on the messages API to ensure requests continue to be blocked using 4.8 at Opus pricing, providing a robust system for API management.

Fable 5 demonstrates significant performance improvements, outperforming models like Opus 4.8, GPT 5.5, and Gemini 3.1 Pro on the Suitebench Pro benchmark, indicating its state-of-the-art capabilities.

> I didn't find something that technically it failed at, so I think these benchmarks are really gonna hold.

## 10:01 - 12:00 Fable 5 demonstrates exceptional vision and document formatting capabilities.

Real-world testing of Fable 5 (also referred to as Mythos 5) revealed a surprising strength in its vision capabilities, particularly for document formatting. The model was evaluated against various documents, with its performance standing out in layout tasks.

For example, when creating a second-grade handwriting sheet, Fable 5 significantly outperformed Opus 4. Fable 5's output featured appropriate spacing, clear readability, and sufficient white space, making it well-suited for a young student. In contrast, Opus 4's version was dense and hard to interpret, even in terms of where to write on the lines.

However, this strength in vision did not extend to prose generation. Fable 5 struggled with producing readable written content, with its output often being described as overly detailed and difficult to parse, akin to an engineer getting too wrapped up in specifics.

> unfortunately it's an engineer, and what's the problem with engineers? They just really get wrapped around the axle on details.

## 12:00 - 13:10 Fable's Overly Detailed and Unreadable Prose for Specifications

The speaker observed that an AI model named Fable generated a markdown document that, despite being long and seemingly intelligent, was exceptionally difficult for humans to parse. The output contained numerous internal references and excessive detail, making it challenging to extract high-level information.

This issue was repeatedly encountered when working with specifications. While Fable's output was undeniably complete, it became nearly impossible for a human to read and comprehend the crucial points due to the overwhelming granularity.

The core problem lies in Fable's tendency to produce very detailed information without providing a way to 'zoom out' or easily see the overall structure and main ideas. This makes it hard to discern the forest from the trees when dealing with complex documents like specs.

A suggestion was made to use models like Sonnet or Opus for generating specifications where human readability is paramount. Fable, conversely, could be better utilized as an orchestrator for execution tasks, where its detailed output is critical for accuracy but doesn't necessarily need to be directly read and interpreted by humans.

> It is just really hard to see the forest for the trees in this particular model.

## 13:10 - 13:30 A skills registry design shows fundamentally terrible results

An attempt to create a skills registry resulted in a "fundamentally terrible design," described as being even worse than typical "AI slop."

The poor design was characterized by basic visual elements, including gray, black, and red colors with simple outlines, indicating a lack of thoughtful or effective design execution.

> it's like fundamentally terrible design.

## 13:30 - 14:20 Fable 5 Underperforms in Frontend Design Tasks

The speaker evaluated Fable 5 for frontend design, specifically attempting to create an impressive skills registry. Despite providing unusually detailed prompts, the model's output was consistently unimpressive. This level of prompting was unnecessary with other models used in the past year for similar frontend tasks.

The design results from Fable 5 were described as legitimately bad, leading to concerns about "design slop." This poor performance was a significant disappointment, suggesting a need to use alternative models like Opus for visual design work.

Fable 5 demonstrated a conservative approach to execution. When tasked with developing the minimum viable product (MVP) of a complex specification intended to deliver customer value, it struggled to produce acceptable design elements.

> I think there's this real balance between design slop and specificity and just shipping like terrible design.

## 14:20 - 16:01 Fable 5 Shows Conservative MVP Execution and Multi-Agent Orchestration Difficulties

Fable 5's Minimum Viable Product was notably narrow and not very useful, potentially due to safeguards built into the model. This conservative approach, common in later Opus models, means users must carefully craft prompts to achieve both ambitious product goals and long-running outcomes from the system.

The speaker attempted to test Claude's dynamic workflows and sub-agent designs for multi-agent capabilities. While some multi-agent runs were successful, the system frequently encountered stalls and errors during orchestration. One instance saw sub-agents halt for three hours, highlighting significant technical challenges.

These orchestration difficulties are believed to stem from issues within Claude Code's implementation rather than the core model itself. For systems promising extensive, long-running prompts and complex tasks, the technical execution of multi-agent workflows is paramount. Despite these challenges, the model is suggested for highly detailed, long-horizon technical problems and vision-related tasks, such as parsing PDFs or generating aesthetically pleasing outputs.

> I also ran into a lot of stalls and errors in using multi-agent orchestration.

## 16:01 - 16:20 Fable 5's Optimal Use Cases and Limitations

Fable 5 demonstrates exceptional performance in specific technical domains. It is particularly effective for problems requiring extreme detail and long-horizon work, such as vision tasks like parsing PDFs or creating well-formatted documents.

However, Fable 5 is not recommended for front-end development, strategy, or specification work. The tool tends to overthink these types of problems, resulting in prose that is described as nearly "imparcel" and often overly complex.

> I definitely wouldn't hand it my front-end work, and I definitely wouldn't hand it strategy or spec work.

## 16:20 - 16:50 Understanding Fable 5's Strengths and Limitations

Fable 5 is not recommended for tasks that require concise and clear output, such as frontend work, strategic planning, or detailed spec writing. The model tends to produce overly dense and difficult-to-parse prose, making it less suitable for these specific applications.

Despite these limitations, Fable 5 is recognized as a valuable tool that has a definite place in a technical stack. It is not dismissed as ineffective overall, but rather its utility is context-dependent.

Users are advised to apply Fable 5 to their most challenging problems, leveraging its strengths where appropriate. Consulting the prompting guide for Fable is suggested to understand how to achieve optimal outcomes and differentiate between its suitable and unsuitable applications.

> It definitely has a place in your stack.

---

Get podcast briefs for shows you follow: https://podbrew.app/