Claude’s Mightiest Model Fable 5 Fully Explained: A Mythos-Level Beast

Anthropic released the new Fable 5 in the wee hours of June 10. It is without a doubt the strongest model on the market right now.

Thought it was Mythos that got released? So what is Fable exactly? Actually, Mythos and Fable run on the same base model, differing only in safety guardrails.

• The original variant is Mythos 5, restricted to governments and designated institutions. Pricing: $25 per million input tokens, $125 per million output tokens.
• Fable 5 is the public-facing version with added guardrails for general use. It is priced at $10 per million input tokens and $50 per million output tokens, which is precisely double the rate of Opus 4.8.

Users of Pro, Max, Team and seat-based Enterprise Plans can use Fable 5 until June 22. It consumes twice the quota of Opus. After that date, usage will be deducted from your usage credits. The API is fully operational today under the model name claude-fable-5.

I've also given it a quick try, and it does perform impressively in practice.

That said, here's the catch. Due to its strict safety guardrails, it declines many tasks and automatically falls back to Opus 4.8 instead.

When using Claude Code, it returns the following message:

Request was blocked
This request triggered safety guardrails. Rephrase your prompt or rewind to continue.

Anthropic claims the block rate is below 5%, but my real-world experience tells me the figure is much higher. So far, I haven't managed to complete a single task fully processed by Fable 5. And I'm not the only one facing this issue.

For instance, a biologist simply sent "hi" to Fable 5, yet the request got blocked. The system flagged it as risky because the model's memory contained relevant biological data. Strangely, the greeting worked just fine when incognito mode was enabled.

I guess I've run too many unusual queries, and the system seems to have me flagged. It really makes me feel like I'm locked out.

For this article, most of the performance-related content is sourced from official data, early tester feedback and Anthropic's official announcement.

Announcement link: anthropic.com/news/claude-fable-5-mythos-5

One Core Model, Two Names

As mentioned earlier, Mythos and Fable are built on the exact same model, differentiated solely by safety guardrails. The version with guardrails is named Fable, while the unrestricted one goes by Mythos.

Etymologically, Mythos derives from Greek and means "myth". Fable comes from the French word Fabula, which refers to "something told". The two terms share similar connotations.

The Mythos tier is a new performance rank created by Anthropic above Opus. Back in April, the company launched the first Mythos-class model, Claude Mythos Preview, via Project Glasswing. It was only available to a select group of cybersecurity defenders and critical software infrastructure providers.

Anthropic stated at the time that once safety guardrails were fully implemented, the Mythos-tier capabilities would be opened to the general public — and that public iteration is Fable.

Here is Anthropic's official model performance ranking:

Official: Mythos > Fable > Opus > Sonnet > Haiku

A True Top-Tier Contender

Following the general rule that the latest flagship models deliver superior performance, Fable 5 and Mythos 5 stand as Anthropic's most powerful offerings to date. Fable 5 achieves top results across nearly all current benchmark tests. Its advantage grows even more prominent when handling long and complex tasks.

For example, Stripe used Fable 5 to migrate a 50-million-line Ruby codebase, completing the entire migration within a single day. The same work would take a human team more than two months to finish manually. Additionally, Fable 5 is more token-efficient than previous Claude models, cutting costs when tackling complex problems. It also scored the highest among frontier models in Cognition's FrontierCode evaluation under moderate reasoning loads.

Early access users have conducted their own tests. Here are some key quotes:

• Michael Truell from Cursor stated it ranks first on CursorBench, unlocking complex long-running tasks that were previously unmanageable.
• Mario Rodriguez of GitHub noted that it delivers greater autonomy and reliability than prior benchmarks when handling intricate, extended coding work.
• Scott Wu from Cognition confirmed it achieved the highest score on FrontierBench, with strong out-of-the-box generalization for unfamiliar tools.
• Michele Catasta of Replit added that it nearly hit full marks in the platform's vibe-coding evaluations.

Suppose lab findings are confirmed.

The capabilities of Mythos 5 in biological research far exceed what standard benchmark tables can capture.

Take drug design as an example. Anthropic's in-house protein design specialists leveraged Mythos 5 to speed up certain stages of the drug development process by roughly ten times. In one case, Mythos 5 worked with protein design and bioinformatics tools autonomously. Its performance matched and even surpassed that of seasoned human professionals. Out of 14 protein targets in the research, it identified 9 promising candidates viable for further drug development.

It generated high-potential candidates for nine targets across immune checkpoints, neurodegenerative diseases and muscular disorders.

Mythos 5 is Anthropic's first model capable of consistently producing innovative and scientifically sound hypotheses. In blind comparative tests, scientists favored Mythos' molecular biology hypotheses over those from Opus-tier models around 80% of the time. Some of its experimental hypotheses have already been validated. For instance, a newly proposed mechanism of a specific protein in E. coli has been confirmed by laboratory research.

In genomics, Mythos 5 independently completed a new study over more than a week of continuous operation. It integrated single-cell data covering millions of cells from 138 animal species, then designed and trained a machine learning model to identify cells with identical functions across distantly related species.

This model outperformed a recently published one in Science, while being 100 times smaller. The research results are set to be published in the coming months.

On Safety Guardrails

I seem to have unusual data stored in the model's memory, so simple tasks can still run successfully for me.

Complex requests, however, get rejected.

Besides drug design, what other applications does Mythos 5 have in biological research?

Anthropic has equipped Fable 5 with a dedicated classifier — a separate AI system that monitors every turn of conversations. It checks if requests fall into three restricted categories: cybersecurity, biochemistry and model distillation. Once a match is detected, the session will automatically switch to Opus 4.8.

Anthropic states the average trigger rate is below 5%, meaning over 95% of conversations remain unaffected. However, this statistic is calculated by total conversation counts, and even a simple "hi" counts as one session. The real-world experience is drastically different when running actual work, especially long tasks.

If a long task hits the classifier at any point during the conversation, the entire session will shift from Fable 5 to Opus 4.8, and all subsequent responses will be handled by Opus 4.8.

Additionally, my extensive chat history sometimes gets flagged for no obvious reason. When the switch occurs, a prompt reading Switched to Opus 4.8 appears at the top of the interface, alongside an option labeled Edit and retry with Fable 5 on the right, allowing users to rephrase prompts and try again.

The phrase "Edit and retry" is quite cleverly worded.

The same applies to Claude Code. When safety guardrails are triggered, it immediately shows "Request was blocked", prompting you to rephrase your query or rewind the conversation.

Now let's break down the three categories of restricted requests for Fable.

1. Cybersecurity

Mythos-tier models excel at vulnerability detection, exploit development and agentic hacking, capable of chaining together tasks like reconnaissance and lateral movement autonomously. The guardrails completely block Fable from making any progress on such work.

Anthropic has conducted extensive red team tests. External bug bounty programs ran for over 1,000 hours without uncovering any universal jailbreaks. External red teams also failed to find workarounds in long-running agentic tasks. Only the UK's AISI managed to make limited progress within a very short initial window.

Tests conducted by an external partner show that Fable 5 has the strictest safety guardrails against malicious cybersecurity requests among all evaluated models, outperforming both Opus 4.8 and 4.7. It blocked all 30 publicly known jailbreak techniques with zero bypasses.

2. Biology and Chemistry

Previously, restrictions only targeted a narrow range of questions related to bioweapons, but Anthropic now deems this insufficient. The team tested Mythos 5 on predicting viral capsid assembly using unpublished candidates from Dyno Therapeutics — a key step in developing AAV gene therapy vectors. The same capabilities could be misused to engineer dangerous viruses.

Mythos-tier models outperform dedicated protein language models purely through biological reasoning. To ensure a safe launch as soon as possible, Fable currently redirects most biochemistry-related requests to Opus 4.8, with restrictions to be gradually relaxed over time.

3. Model Distillation

Anthropic has detected large-scale distillation attempts in the past. Any such activity will now be flagged by the classifier and routed to Opus 4.8.

Another change directly impacts enterprise customers. All traffic for Mythos-tier models, whether via Anthropic's native interface or third-party platforms, is subject to a 30-day data retention policy. The stored data will not be used for model training or any purposes other than safety monitoring, and all records of human access will be deleted after 30 days.

Free Access Until June 22

A pop-up notice on the web version informs users that Fable 5 can be used within existing plan quotas before June 22, though it consumes twice the quota of Opus. Starting June 23, it will no longer be included in standard subscriptions, and usage will require usage credits. It will be reinstated as a standard subscription feature once system capacity allows.

It consumes twice the quota of Opus. Starting June 23, you will need to spend your own usage credits to use it.

Fable 5 is directly selectable in the model list on the web interface, marked with Included until June 22. By contrast, it does not appear in the default list in Claude Code, whether in the terminal or the Claude app. You have to manually enter /model claude-fable-5 to switch to it.

The claude-fable-5 model is available across multiple platforms, including the official Claude API, AWS Claude Platform, Amazon Bedrock, Vertex AI and Microsoft Foundry.

Mythos 5 remains private. It is exclusively offered to Project Glasswing partners (with cybersecurity guardrails lifted) and a selected group of biologists who will gain access soon (with biochemistry guardrails removed). Anthropic states it will gradually expand the trusted access list for Mythos 5 in collaboration with the U.S. government. Dedicated application channels will be launched separately for cybersecurity organizations and biological research institutions.