# Anthropic Built an AI Too Dangerous to Release. Then They Released It Anyway.

By Manish Parasher (@manishparasher) · Published 2026-06-11

Canonical: https://voce.com/@manishparasher/anthropic-built-dangerous-release-released-anyway-dekpok

---

In April 2026, Anthropic's most powerful AI model quietly found thousands of zero-day software vulnerabilities on its own — including bugs that had been sitting undetected in OpenBSD for 27 years.

Anthropic's response? Lock it away. The model — Claude Mythos — was deemed too capable for general release and shipped only to a handful of vetted cybersecurity firms and infrastructure providers through a restricted program called Project Glasswing.

Then, on June 9, 2026 — three days ago — Anthropic did something unexpected.

They handed a version of it to the public.

That version is called **Claude Fable 5**. And it changes the benchmark for what an AI model can actually do.

## What Fable 5 Actually Is

Let's be precise about this, because the naming is confusing.

On June 9, 2026, Anthropic released two models built on the same underlying weights: Claude Fable 5 and Claude Mythos 5. Same model. Different guardrails.

Fable 5 is the public version — wrapped in safety classifiers that intercept high-risk queries. Mythos 5 is the same thing with those classifiers partially lifted, available only to approved organizations through Project Glasswing.

Anthropic describes the new "Mythos-class" tier as exceeding the capability of anything it has previously released for general use.

That's not marketing language. The benchmarks back it up.

## The Numbers That Actually Matter

Benchmark scores get inflated and misused constantly in AI marketing. So here's the honest version of what Fable 5's numbers mean.

**SWE-Bench Pro: 80.3%**

Fable 5 posts 80.3% on SWE-Bench Pro versus Opus 4.8 at 69.2% and GPT-5.5 at 58.6%. This benchmark tests whether a model can resolve genuinely hard software engineering problems pulled from real repositories — not toy examples. An 11-point jump over Anthropic's own previous best is a real capability gap, not benchmark tuning.

**FrontierCode Diamond: 29.3%**

FrontierCode Diamond — which tests production-quality coding — came in at 29.3%, more than double Opus 4.8's 13.4%, with GPT-5.5 at 5.7%. More than double. On a benchmark designed to catch the kind of mistakes that actually break production code.

**The real-world proof point:**

Fable 5 finished a migration in a 50-million-line codebase in a single day. That is not a benchmark. That is Stripe's actual testing. And Andrej Karpathy — one of the most respected voices in AI — called it "a major-version-bump-deserving step change forward."

When a researcher of that calibre uses language like that, it's worth paying attention.

## The Part Nobody Wants to Talk About: The Safety Fallback

Here's the most interesting — and most overlooked — detail in the entire Fable 5 launch.

In high-risk areas like cybersecurity, biology, chemistry, and distillation, Fable 5 blocks responses and falls back to Claude Opus 4.8.

Read that again. You can pay for a Mythos-class model, send it a query, and silently receive an Opus 4.8 response instead.

For most developers, this is invisible. The work that benefits most from Fable 5 — large-scale software engineering, complex knowledge work, long-horizon agentic tasks — sits well outside those guarded categories.

But for teams doing legitimate security research or life-sciences work, this matters. The fallback doesn't announce itself. You have to know it exists.

Its strongest quality is the ability to keep making progress when a task becomes long, unclear, and full of failed attempts. That's where Fable 5 separates itself. Not on clean, well-specified tasks. On the messy, multi-hour, multi-step ones where previous models would stall.

## The Pricing Reality Check

Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens — double the rate of Claude Opus 4.8 ($5/$25) but less than half the price of the earlier Mythos Preview ($25/$125).

That pricing trajectory is the real story. Mythos Preview was priced at a level that made it inaccessible to most developers. Fable 5 cuts that cost by more than half while delivering the same underlying capability with guardrails.

For Claude.ai subscribers: Fable 5 counts as 2x usage on subscription plans. It's free on Pro, Max, Team, and Enterprise through June 22, 2026 — after that it requires usage credits until capacity allows it to return as a standard feature.

If you're on a subscription plan, this week is the window to test it at no extra cost.

## Who Should Actually Care

**Use Fable 5 if you're:**

-   Running large-scale autonomous coding tasks that take hours, not minutes
    
-   Working on complex document reasoning, legal analysis, or financial modelling
    
-   Building agentic workflows where the model needs to self-correct across many steps
    
-   Hitting the ceiling of what Opus 4.8 can sustain over long contexts
    

**Stick with Opus 4.8 if you're:**

-   Doing routine professional work — writing, summarization, research, Q&A
    
-   Working in cybersecurity, biology, or chemistry (where Fable 5 silently falls back anyway)
    
-   Cost-sensitive and running high message volumes
    

Opus 4.8 remains the more practical choice for routine professional work. Fable 5 becomes more interesting as the task grows.

The model ID for the API is `claude-fable-5`. Context window is 1 million input tokens with 128K output. Extended thinking is supported.

## Why This Moment Is Bigger Than One Model Release

Fable 5 isn't just a faster Claude. It's a signal about where the capability ceiling is moving.

The April 2026 Mythos Preview found thousands of zero-day vulnerabilities autonomously, including 27-year-old OpenBSD bugs — and was deemed too powerful for general release. That's the underlying model underneath what you can now access.

The architecture of Fable 5 — same weights, safety classifiers on top — is Anthropic's answer to a question every frontier lab is wrestling with: how do you release your most capable model without releasing your most dangerous one?

Whether the classifier approach holds up under pressure from sophisticated users is a real open question. But the fact that Anthropic chose to release at all — after months of locking Mythos away — says something about where they think the risk-benefit calculation now lands.

The most powerful publicly available Claude model in history dropped three days ago. Most people are still using the one from last month.

**Have you tried Fable 5 yet? And if you have — what task did you run it on first? Drop it in the comments. Genuinely curious what people are testing.**