Fable 5 is back: what it is, why Anthropic pulled it, and what changed

On June 9, 2026, Anthropic released Claude Fable 5, its most capable model to date. Three days later it pulled it globally. On July 1 it became available again. Here's what it is, why this happened, and why the cybersecurity angle behind the whole episode matters to us in particular.

What Claude Fable 5 is

Fable 5 is the general-use version of a "Mythos"-class model, the highest capability tier Anthropic has built to date. It's state of the art on nearly every public benchmark: software engineering, analysis and reasoning, vision, and long context (it works well even with millions of tokens in a single conversation). Running inside an agent harness like Claude Code, it can sustain tasks for days: planning across stages, delegating to sub-agents, and checking its own work.

Stripe, one of the first to test it, reported it compressed months of engineering work into days. That's the kind of capability jump we're talking about, not an incremental version bump.

Why it was pulled

Days after launch, Amazon researchers found a way around the model's safeguards: they got Fable 5 to identify software vulnerabilities and, in one case, to generate code showing how to exploit one.

The US government responded with an export control order on June 12, requiring that access be restricted for foreign users. The problem for Anthropic was technical: there was no reliable way to verify each user's nationality in real time. Facing that constraint, the company chose to suspend access to both Fable 5 and Mythos 5 for everyone, not just the flagged cases.

What changed for the relaunch

Anthropic trained a new safety classifier that blocks the reported behavior in over 99% of cases. It added this to a "defense in depth" setup: several layers of smaller classifiers running in parallel, detecting potentially risky cybersecurity tasks and routing them to Claude Opus 4.8 instead of letting them through Fable 5.

With those safeguards in place, the government lifted the export control on June 30, and access was restored on July 1 across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork.

The part we care about most: Mythos 5 and cybersecurity

Fable 5 isn't the only model in this family. Mythos 5 shares the same base model but without most of Fable's safeguards, and it isn't available to the general public: it's limited to authorized defensive cybersecurity research partners, under a specific Anthropic program (Project Glasswing).

In other words: the same company that had to block its most powerful model because someone used it to find software vulnerabilities also keeps, in parallel, an unrestricted version exclusively so cybersecurity teams can use it defensively, under controlled access. The Fable 5 incident itself is the proof of concept: a model at this level can find real security flaws in real code, for better or worse, depending on who operates it and under what safeguards.

That's exactly the space we've been watching at Manivela as our next line of work: cybersecurity applied to organizations that currently have no active auditing or defense at all (municipalities, small businesses, public agencies). AI doesn't replace a security team, but a model capable of reasoning about code and spotting vulnerabilities is a real tool for closing that gap, as long as it's used with the same defense-in-depth mindset Anthropic applied here: layered review, controlled access, and a human always validating before anything gets acted on.

What this means in practice

This episode confirms something we've been pointing out for a while: AI models move fast enough that safeguards sometimes arrive after the fact. That's not a reason to write off the tool, but it is a reason not to depend on a single model, a single provider, or blind trust in safeguards that are still being tested in the real world.

At Manivela we pick the model for the task for exactly this reason. A frontier model like Fable 5 makes sense for complex reasoning or to speed up development, but every line it produces goes through human review before it reaches production.

In short

Fable 5 came back online after three weeks blocked over a real security issue, not a business decision. And the fact that Anthropic keeps an unrestricted version exclusively for defensive cybersecurity confirms this is a field that's going to grow fast. If your organization needs to start thinking about its digital security, or how to adopt AI responsibly, let's talk.