---
title: "Claude Fable 5 and Mythos 5 — Anthropic's most powerful model, with an asterisk"
description: "Claude Fable 5 and Mythos 5 are Anthropic's new model class above Opus. Benchmarks, pricing, safeguards, data retention, and what changes on June 23, 2026."
date: 2026-06-09
category: AI
tags: ["AI", "Anthropic", "Claude", "LLM", "Security"]
url: https://uper.pl/en/blog/claude-fable-5/
---

# Claude Fable 5 and Mythos 5 — Anthropic's most powerful model goes public (with an asterisk)

On June 9, 2026, Anthropic did something that top-tier AI companies had so far avoided. It deliberately released a Mythos-class model for general use — the highest capability tier in the entire Claude lineup, sitting above the Opus family. The new model is called Claude Fable 5. Alongside it comes a twin, more powerful version, Claude Mythos 5, available only to a narrow group of trusted partners.

For the first time, a company of this stature is saying openly that it has a model which, without safeguards, would be too dangerous to simply ship. And it releases it publicly anyway, wrapping it in a layer of cautious protective mechanisms.

## Fable 5 vs. Mythos 5 — same engine, different brakes

Fable 5 and Mythos 5 are the exact same base model. The only thing that separates them is the set of safeguards.

- **Mythos 5** is the "raw" version, with some restrictions lifted. It has the strongest cybersecurity capabilities in the world and goes only to vetted entities under Project Glasswing. Initially to partners working with the US government, soon also to selected biology researchers.
- **Fable 5** is the same intelligence, but "insured" — prepared so that it can be safely made available to everyone.

The name itself hints at this. "Fable" comes from the Latin *fabula*, "that which is told," and is close to the Greek *mythos*. The models are distinguished solely by their safeguards, which is why they were given separate names.

Anthropic took a similar step earlier, when [the Capybara class and Claude Mythos Preview](/en/blog/claude-mythos-capybara-class/) defined a new model tier above Opus.

## What Fable 5 can actually do

Anthropic doesn't mince words: Fable 5's capabilities exceed everything the company has previously made publicly available. The model is state-of-the-art on nearly every benchmark tested. A few numbers that make an impression:

| Area | Benchmark | Fable / Mythos 5 | Opus 4.8 | GPT 5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|---|
| Agentic coding | SWE-Bench Pro | **80.3%** | 69.2% | 58.6% | 54.2% |
| Agentic coding | Terminal-Bench 2.1 | **88.0%** | 82.7% | 83.4% | 70.7% |
| Frontier coding | FrontierCode (Diamond) | **29.3%** | 13.4% | 5.7% | — |
| Knowledge work | GDPval-AA | **1932** | 1890 | 1769 | 1314 |
| Vision | GDP.pdf (no tools) | **29.8%** | 22.5% | 24.9% | 16.7% |
| Multidisciplinary reasoning | Humanity's Last Exam (with tools) | 64.5% | 57.9% | 52.2% | 51.4% |
| Computer use | OSWorld-Verified | 85.0% | 83.4% | 78.7% | 76.2% |

The lead is clearest in coding and analytical work, and it grows as the task gets more complex.

The most vivid example comes from Stripe, which tested the model before launch. In a 50-million-line Ruby codebase, Fable 5 performed a codebase-wide migration in a single day. Doing it by hand would have taken a whole team over two months. The model is also more token-efficient than its predecessors.

The second strong suit is vision. Fable 5 can read precise values off detailed scientific charts, and even reconstruct a web app's source code from screenshots alone. One small detail captures it well: earlier models couldn't get through Pokémon FireRed even with elaborate "harnesses," while Fable 5 finished the game using nothing but raw frames of the screen.

## The core advantage: working autonomously for hours

If there's one sentence to take away from the announcement, it's this: the longer and more complex the task, the bigger Fable 5's lead over other models.

The model holds focus across millions of tokens on long-running tasks and improves its own output using notes it writes for itself. In a test with the deck-building game *Slay the Spire*, access to persistent file-based memory boosted its performance three times more than it did for Opus 4.8. For the user, this means a model you can hand an ambitious, multi-stage project and expect sensible progress without constant micromanagement.

## Safety as the price of admission

This is the part worth knowing before you get carried away. Anthropic openly admits that a model of this class carries serious risk. Without safeguards, its cybersecurity capabilities could be used to cause real damage. That's why Fable 5 ships with a layer of classifiers — separate AI systems that detect potential misuse and attempts to bypass the safeguards.

Here's how it works. When a query touches on cybersecurity, biology and chemistry, or model distillation attempts, Fable 5 doesn't answer; the weaker (though still capable) Claude Opus 4.8 takes over instead. The user is informed whenever this happens.

And this is where the first downsides appear.

**The first is overcautious safeguards.** Anthropic deliberately tuned them conservatively, so they sometimes catch entirely harmless queries. According to the company, the fallback triggers on average in fewer than 5% of sessions, meaning over 95% of sessions run fully on Fable 5. For some users, especially those working near sensitive topics, it can still be frustrating. Anthropic says it will narrow these safeguards after launch.

**The second is mandatory data retention.** For Fable 5, Mythos 5, and future models of this class, all traffic — inputs and outputs — is retained for 30 days, solely for safety purposes: detecting new attacks and reducing false positives. The data is not used to train models or for any non-safety purpose, and the company has added new privacy protections. For organizations sensitive to data handling, it nonetheless remains a real constraint.

## Availability and pricing — where you need to watch out most

This is the section where the enthusiasm should be weighed against your wallet.

API pricing is high: **$10 per million input tokens and $50 per million output tokens**. That's admittedly less than half the price of the earlier Mythos Preview, and there are discounts for batch processing and caching. Still, on autonomous tasks that run for hours and burn through enormous numbers of tokens, the bill can climb fast. The model runs as `claude-fable-5` via the API.

The biggest catch concerns subscriptions. Anthropic expects very high and hard-to-predict demand, so the rollout on subscription plans is staged:

- From launch **through June 22**, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost.
- **On June 23**, the model will be removed from those plans. Continued use will require usage credits, i.e. pay-as-you-go. If capacity allows, the included window will be extended.
- Eventually, once the infrastructure permits, Anthropic wants to restore Fable 5 as a standard part of subscriptions. It says it will do so as quickly as it can.

So if you want to test the model "for free" within your subscription, you realistically have until June 22, 2026. After that, at least for a while, you pay for what you use.

And here's the downside that will hit the largest group of users. A good chunk of developers work day to day on subscriptions (Pro, Max, Team) rather than on the per-token API. For them, continuing to use Fable 5 after June 23 means reaching for an additional, paid mechanism: usage credits or the API directly. That's a cost on top of what they already pay in their subscription. Until Anthropic restores the model as a standard subscription feature, for many people this will be a barrier, not just an inconvenience.

There's a third limitation, too. Mythos 5, the fully unlocked version, stays out of reach for ordinary users. It goes only to Glasswing partners and, soon, a handful of biology researchers. The public gets only the version with the brakes on.

## The balance sheet: pros vs. cons

**On the plus side:**

- Top scores on nearly every benchmark, with a wide lead in coding and analysis.
- Genuine autonomous work for hours. The model plans, tests its own output, and handles ambiguous tasks.
- Strong vision and long-term memory.
- A transparent safety model: instead of a hard refusal, you get a response from Opus 4.8.
- A lower price than the earlier Mythos Preview.

**On the minus side:**

- High API costs ($10/$50 per million tokens), which rack up quickly on long tasks.
- Free access on subscriptions only until June 22, then usage credits.
- For most subscription users, having to buy API access or credits will be a real barrier.
- The conservative safeguards can be overzealous and reroute some harmless queries to the weaker model.
- Mandatory 30-day data retention.
- The most powerful version (Mythos 5) is unavailable to the general public.

## Who it makes sense for

Fable 5 shows its claws most where other models needed constant human supervision. It's worth testing for:

- long, complex coding projects: migrations, refactors, tasks spread across many stages;
- building autonomous agents that you hand goals to rather than single commands;
- advanced analysis of documents, charts, and tables (finance, research, enterprise workflows);
- vision tasks where you need to read data off images precisely.

If, on the other hand, your work regularly brushes up against cybersecurity or biology, expect some queries to fall back to Opus 4.8. In that case it's worth considering whether these limitations and the data retention policy suit you at all.

## Summary

Claude Fable 5 isn't just a "more powerful Claude." It's a new level of agentic intelligence, designed for tasks that until now needed a human at the helm. The whole launch is also an exercise in balance. Anthropic shows that a system this advanced can be released responsibly, but it does so carefully: with brakes, with mandatory data retention, with a short free-access window, and with the most powerful version kept away from the general public.

For developers and knowledge-work teams, it's nonetheless real access to a model capable of multi-day autonomous work — something that until recently was the domain of closed experiments. The coming weeks, and especially what happens with pricing after June 23, will show how much it changes everyday work. For now, it's worth testing while it's still part of the subscription.

## Sources

1. **Official Anthropic announcement — Claude Fable 5 and Claude Mythos 5**
[https://www.anthropic.com/news/claude-fable-5-mythos-5](https://www.anthropic.com/news/claude-fable-5-mythos-5)
2. **Fable model page**
[https://www.anthropic.com/claude/fable](https://www.anthropic.com/claude/fable)
3. **Mythos model page**
[https://www.anthropic.com/claude/mythos](https://www.anthropic.com/claude/mythos)
4. **Project Glasswing**
[https://www.anthropic.com/glasswing](https://www.anthropic.com/glasswing)
5. **Model system card**
[https://anthropic.com/claude-fable-5-mythos-5-system-card](https://anthropic.com/claude-fable-5-mythos-5-system-card)
6. **Pricing and API availability**
[https://platform.claude.com/docs/en/about-claude/models/overview](https://platform.claude.com/docs/en/about-claude/models/overview)
7. **Managing usage credits**
[https://support.claude.com/en/articles/12429409-manage-usage-credits-for-paid-claude-plans](https://support.claude.com/en/articles/12429409-manage-usage-credits-for-paid-claude-plans)