What Is Claude Fable 5? Inside Anthropic’s Mythos-Class Top Model

Since arriving in June 2026, Anthropic’s top model Claude Fable 5 has dominated conversation in AI circles: an unusual withdrawal and return, a limited window of access inside subscriptions, and a widespread verdict that it is a class apart. Here is what it is, drawing on official information, community reaction, and my own heavy use of it.

What Fable 5 is

Fable 5 is the top model Anthropic made generally available on 9 June 2026. The significant part is that it is the first time the Mythos-class tier — held back from general release as too capable — was opened to general users, with safety mechanisms in place.

On specifications, it handles a one-million-token context window and output up to 128,000 tokens, and it is aimed at complex reasoning and long-running autonomous agent work. After release it was temporarily withdrawn following a legal directive from the US government, then restored — an unusual path that added to the attention it received.

The reaction: capability praised, operation questioned

Assessments of capability have been consistently high, with top-tier placement on composite benchmarks and particular praise for complex problem-solving and for reading intent.

The community has also raised the safety mechanisms firing too readily, and concerns about pricing — the shifting periods and allowances within subscriptions, and the cost of metered use. Access inside subscriptions has been limited-period and capped, repeatedly extended, which makes how long and how much you can use it hard to read. That uncertainty is a shared frustration.

My take: the revolution is that it leaves a system behind

I subscribed to Claude’s top tier for the first time during this window and used the allowance until it ran out. Three impressions:

• Reading between the lines is in a different class. Throw a vague idea at it in vague words and it takes the intent and gives it shape. As a partner for going from nothing to something, it is about as good as it gets
• Problems you gave up on get solved. Hand it material on something no AI had cracked before and it goes through. The right way to test a top model is to throw your list of past failures at it
• It leaves a system behind. Once Fable 5 has built the structure, ordinary models such as Opus 4.8 can run it. Spending the window on assets beats spending it on one-off output

The full retrospective is in using up Fable 5, and the list of what I actually built is in what I built with Fable 5.

Not universal: there is an element of chance

Unqualified praise would not be fair, so the weaknesses. Fable 5 is less faithful to instructions than inclined to run with what it thinks you want, and a vague instruction sends it off on its own interpretation. It is capable enough to reach a reasonable destination, but whether that is the destination you wanted is partly luck — an impression that matches the wider reaction. For work where the task is already clear, a model that follows instructions more literally, such as GPT-5.6, is the better choice.

Worth trying if you get the chance

Fable 5 is the best available way to feel how far the frontier has moved. It is overspecified for daily work, but pointed at unresolved problems and at building systems, it returns more than it costs. If access opens inside a subscription again, it is worth experiencing once.

For the Claude service itself, and for underlying terms such as context window and AI agent, see the AI glossary on this site.

Availability and pricing information in this article is as of 17 July 2026. Check Anthropic’s own announcements for current details.