Anthropic's most powerful general-use model, Claude Fable 5, comes back online worldwide today after the US government pulled it offline three weeks ago. The whole episode is confusing if you don't know what Fable actually is, why it sits in its own class, or why a government would care about a chatbot. This is a plain-English guide to what Fable is, how Claude's model tiers work, why it got shut down and switched back on, and what it's genuinely good at (and where it's the wrong tool). At the end, a short brief on Claude Sonnet 5, which also launched this week and matters more for everyday work.

First, what a "model tier" even means

When people say "Claude" or "ChatGPT," they're usually talking about a single product. Under the hood, though, each company offers a range of models at different levels of power, speed, and price. You pick the one that fits the job, the same way you'd pick a compact car for the school run and a truck for a move.

Claude's lineup runs roughly from lightest to heaviest:

The practical point is that a "more powerful" model is not automatically the right choice. Higher tiers cost more and are slower, so the skill is matching the model to the task. Most business work never needs to be at the top of the ladder.

So what is Claude Fable, exactly?

In April, Anthropic introduced its first Mythos-class model to a tiny group of vetted cybersecurity defenders under a program called Project Glasswing. It was too capable and too easy to misuse to hand to the public. The obvious question followed: could that level of capability ever be made safe enough for general release?

Fable 5 is Anthropic's answer. On June 9, the company launched two models built on the same underlying brain:

• Claude Mythos 5 is the raw model with its guardrails largely removed. It's restricted to a small set of trusted cyber defenders and, soon, vetted biology researchers. Anthropic calls it the strongest cybersecurity model in the world.
• Claude Fable 5 is that same model wrapped in the strongest safety controls Anthropic has ever shipped, so it can be made available to everyone.

The names capture the distinction. "Mythos" and "Fable" both mean roughly "a story that is told." The only real difference between the two models is the safeguards.

That's the headline: Fable 5 is the most capable model Anthropic has ever released to the general public. Before Fable, that spot belonged to Opus.

Why Fable is different from the Claude you already use

Two things set it apart.

It's built for long, unsupervised work. Most models are strongest in a quick back-and-forth: you ask, it answers. Fable is built to run for hours, even days, on a single complex assignment. It plans across stages, hands subtasks to helper agents, writes its own tests to check itself, and uses vision to compare its output against the goal. Anthropic's framing is that you hand off a large project and review the finished work, rather than supervising every step. As a headline example, Stripe reported that Fable did a company-wide code migration on a 50-million-line codebase in a day, work that would have taken a team over two months by hand.

It comes with unusually aggressive safety guardrails. Because the underlying model is powerful enough to be genuinely dangerous in areas like cybersecurity and biology, Fable ships with automated filters (which Anthropic calls classifiers) that monitor for risky requests. When one trips, your request doesn't get refused outright. It quietly gets handled by Opus 4.8, the next model down, and you're told it happened. Anthropic deliberately tuned these filters to be cautious, so they sometimes catch harmless requests too. In practice, more than 95% of sessions never hit a filter at all. Using Fable also requires 30-day data retention for safety monitoring, a detail that matters if you work with sensitive or regulated data.

Why the government shut it down

On Friday, June 12, at 5:21pm ET, the US government sent Anthropic an export-control directive citing national security. It ordered access to Fable 5 and Mythos 5 blocked for any foreign national, inside or outside the US. Because Anthropic had no way to verify every user's nationality in real time, the only way to comply was to shut both models off for everyone. All other Claude models kept running.

The trigger: researchers at Amazon found a way to bypass Fable's safeguards, prompting it to identify a set of software vulnerabilities and, in one case, to produce code showing how a vulnerability could be exploited. That's the kind of finding that sets off alarms about AI making cyberattacks easier.

Anthropic's counter-argument, which the following weeks largely bore out, was that this was a borderline case rather than a genuine danger. Its testing showed that far less powerful models, including Opus 4.8, OpenAI's GPT-5.5, and even the lightweight Claude Haiku 4.5, could find the same vulnerabilities and reproduce the same demonstration. The behavior amounted to routine defensive security work that professionals perform every day and exposed no capability unique to the Mythos-class model. In other words, the guardrail got nudged open slightly, but not onto anything you couldn't already get elsewhere.

Why it's back today

The export controls were lifted on June 30, and Fable 5 returns to users worldwide starting today, July 1, across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork. On Pro, Max, Team, and select Enterprise plans, it's included for up to half of your weekly usage limits through July 7, after which continued use is charged to usage credits. Availability on AWS, Google Cloud, and Microsoft Foundry is being restored as fast as Anthropic can manage. Mythos 5 was separately switched back on for a set of vetted US organizations on June 26.

Three things came out of the episode that are more interesting than the pause itself:

1. A tighter fix. Anthropic trained a new filter that blocks the specific technique from the Amazon report in over 99% of cases. Government testers at the Center for AI Standards and Innovation reviewed the old and new safeguards and agreed they're extremely strong. The trade-off: the stricter filter now flags more harmless coding and debugging requests, which Anthropic says it will continue to tune.

2. An industry-wide standard for rating jailbreaks. Today, there's no shared way to say how serious a given "jailbreak" (a trick that bypasses a model's safety controls) actually is, which is part of why this got messy. Anthropic, together with Amazon, Microsoft, Google, and other partners, is proposing a common scoring framework based on four questions: how much new capability the trick unlocks, how many different attacks it enables, how easy it is to turn into a real attack, and how easy it is to discover. A shared yardstick lets developers and governments respond proportionally rather than hitting the kill switch every time.

3. Deeper government coordination. Anthropic committed to giving government partners early access to frontier models and their safeguards for independent testing before broad release, faster information sharing on new jailbreaks, and joint research. The bigger picture is an attempt to build a predictable, rules-based process for releasing powerful models, rather than the abrupt on-off switch everyone just lived through.

What Fable is genuinely good at

This is where it earns its place. Fable is the model to reach for when the work is big, complex, and would otherwise eat a large chunk of skilled human time:

• Ambitious, long-running coding. Large migrations, complex builds, and multi-hour or multi-day autonomous sessions where it plans, writes its own tests, and checks its work. Early testers (Cursor, GitHub, Cognition, Replit) called it a clear step up on exactly this kind of long-horizon work.
• Deep knowledge work you hand off and review. Multi-stage research, analysis, and drafting deliverables with minimal supervision. It topped senior-level finance and analytics benchmarks, and legal testers said its contract redlines matched or beat their existing model.
• Document-heavy and visual work. It reads charts, tables, and diagrams buried in PDFs, which is valuable in finance, legal, analytics, and architecture.
• Frontier scientific and technical research. Testers reported research-scientist-grade reasoning, in one physics case reaching in 36 hours what a rival model took four days to do.

Fable shines the longer and more complex the task. On short, simple questions, its edge over cheaper models mostly disappears.

Where Fable is the wrong tool

Being the most powerful option doesn't make it the right default. A few honest limits, starting with price:

• It's expensive. Fable runs several times the price of Opus and many times the price of Sonnet. For routine drafting, summarizing, and Q&A, you're paying a steep premium for capability you won't use.
• It's overkill for everyday tasks. For most day-to-day work, Sonnet or Opus will be faster, cheaper, and plenty capable. Save Fable for the genuinely hard, long-running jobs.
• It will refuse or hand off some requests. Anything the safeguards flag as cybersecurity, biology, or chemistry gets routed to Opus 4.8 instead, and the newly tightened filter now flags more benign coding and debugging requests as false positives. If your work lives near those areas, expect occasional friction.
• It requires 30-day data retention. For safety monitoring, Fable traffic is retained for 30 days. Anthropic says it won't be used for training, but if you're in a regulated field, factor that in before pushing sensitive data through it.

The short version

Fable 5 is Anthropic's top-tier public model: a version of its most powerful "Mythos-class" system, made safe enough for general release with heavy guardrails. It got pulled offline over a cybersecurity scare that turned out to be a borderline false alarm, and it's back today with a tighter fix and a new cross-industry safety standard. It's outstanding for big, long-running, complex work and a poor, pricey fit for everyday tasks. For those, there's a better new option.

The one most businesses will actually use: Claude Sonnet 5

While Fable grabbed the headlines, the more practical release for most businesses is Claude Sonnet 5, out June 30. Sonnet is the everyday workhorse tier, and this is the most capable and most "agentic" version yet, meaning it's markedly better at planning, using tools like browsers and terminals, and finishing multi-step tasks on its own without stopping halfway.

The notable part is the price-to-power ratio: Sonnet 5 performs nearly as well as Opus 4.8 on many tasks, at a fraction of the cost. Launch pricing is $2 per million words of input and $10 per million of output through August 31, then $3 and $15, versus Opus at $5 and $25. Testers described it finishing complex jobs that older Sonnet models would abandon, and writing its own tests to verify a fix without being asked. It's available everywhere today, and it's the default model on Free and Pro plans, so you may already be using it.

Why it matters: For the vast majority of real work, Sonnet 5 is the sensible default: near-top-tier capability at workhorse prices. Reserve Fable for the rare, genuinely hard project that justifies the cost, and let Sonnet handle the rest.

Keep reading