Claude Fable 5: Inside Anthropic's First Public Mythos-Class Model

What Anthropic Actually Released

Anthropic launched Claude Fable 5 on the morning of 9 June 2026, and it is a meaningfully bigger deal than a routine version bump. Fable 5 is the first publicly available model from a new top tier — what Anthropic calls “Mythos-class” — that sits above the familiar Opus line. The company describes it plainly as the most capable model it has ever made generally available.

It launched alongside a sibling, Claude Mythos 5. The two share the same underlying model. The difference is access and guardrails: Fable 5 is the public version, wrapped in safety classifiers, while Mythos 5 has some of those classifiers lifted and is restricted to vetted partners. Think of it less as a “small versus large” pairing and more as the same engine sold with two very different sets of brakes.

The headline specifications: a 1 million-token context window, up to 128,000 output tokens per request, and a knowledge cut-off of January 2026. It is available immediately via the Claude API (model ID claude-fable-5), in Claude Code, across the Claude apps, and through AWS (Bedrock and the Claude Platform), Google Cloud's Vertex AI, Microsoft Foundry and GitHub Copilot. For a refresher on how these models differ under the hood, the AI Fundamentals course is a good starting point.

The Mythos Backstory

To understand why Fable 5 matters, you need the story that led to it. In April 2026, Anthropic announced a model it called Mythos and described as a step change in capability — then explicitly declined to release it. The stated reason was raw offensive-security ability. In testing, the model could autonomously find and exploit software vulnerabilities at a scale earlier models could not approach: by some accounts it surfaced thousands of zero-day vulnerabilities and produced 181 working Firefox exploits on a setup where the previous Claude generation managed two.

Rather than ship it, Anthropic stood up Project Glasswing — a consortium of around 50 major technology and infrastructure organisations, reportedly including Amazon, Google, Microsoft, Apple, Cisco and JPMorgan Chase, plus the US government — to use the model's full power for defensive security while a guardrailed public version was built. Fable 5 is that public version. Anthropic frames the launch as honouring its stated goal of eventually deploying Mythos-class models at scale, now that it believes the safeguards are ready.

What Fable 5 Can Do

The capability story has two halves: the benchmark numbers, and what early users are reporting in the wild. The benchmarks are lopsided in Fable's favour — with the important caveat that, except where independently confirmed, these are Anthropic-reported figures.

Benchmarks

The number drawing the most attention is agentic coding. On SWE-Bench Pro — real software-engineering tasks drawn from public repositories — Fable 5 scores roughly 11 points clear of the next best model. That gap is larger than the gap between Opus 4.8 and Gemini 3.1 Pro.

One detail worth reading carefully: on a handful of starred benchmarks — notably cybersecurity and biology — the figure Anthropic publishes is actually Mythos 5's, not the public Fable 5's. Because Fable routes those exact topics to Opus 4.8, the model you can actually use may score lower there than the table suggests. The honest framing, which the better coverage has adopted, is that Fable 5 leads clearly on the work most businesses actually do, and that its lead widens the longer and harder the task gets.

What early users are seeing

The qualitative reports are more persuasive than any single score, because they all point the same way — this is a model built for long, autonomous work:

• The Stripe migration: the most-quoted data point of the launch. In a 50-million-line Ruby codebase, Fable 5 performed a codebase-wide migration in a single day that the team estimated would have taken more than two months by hand.
• Endurance with memory: given access to persistent, file-based memory, Fable reached the final act of the game Slay the Spire three times more often than Opus 4.8 with the same upgrade — a clean illustration of how much better it holds a thread across a very long task.
• Less scaffolding for computer use: it completed Pokémon FireRed with a minimal, vision-only harness, where earlier Claude models needed elaborate helper tooling to get close.
• Independent hands-on: respected developer Simon Willison, who did not have early access, spent an afternoon with it and called it “something of a beast” — slow and expensive, but able to churn through almost everything he threw at it. In a single day he had it write an almost-complete release of his open-source library, work he reckoned felt like several days' worth.

That last point comes with a price tag worth internalising: Willison burned through more than $110 of tokens in that one day. Fable is genuinely token-hungry, which matters for how you use it. If you want to put these long-horizon patterns to work in your own stack, the AI Agents course covers building and supervising agentic workflows in detail.

The Safety Architecture (the genuinely new part)

This is the most novel thing about Fable 5, and it changes how the model behaves day to day. Fable does not refuse risky requests the way models usually do. Instead, a separate set of classifiers watches every session for requests in three high-risk areas — offensive cybersecurity, biology and chemistry, and model distillation. When one triggers, the request is quietly handed to Claude Opus 4.8 instead, and you are told the handoff happened.

Anthropic's argument is that a fallback to Opus — itself a highly capable model — is a far better experience than a flat refusal. The company says more than 95% of sessions involve no fallback at all, and for those, Fable performs essentially identically to Mythos 5. There is even new billing plumbing for this: a blocked request returns a credit token so the retry on Opus is billed as though the whole conversation had been on Opus from the start, with the switching cost refunded.

The design goal is “degrade, don't refuse.” Whether that feels reassuring or unsettling depends a lot on what you were trying to do when the handoff fired.

The Controversies

This is where the past 24 hours of online discussion has concentrated, and it is the part most launch coverage glosses over. There are three live debates.

1. The safety filters are over-sensitive

Anthropic admits the classifiers are deliberately conservative — tuned first for robustness, which means benign technical work sometimes trips them, and that the biology and chemistry coverage is narrower than ideal because the company prioritised shipping speed over precision. Users surfaced the consequences within hours: reports circulated of the word “cancer” being flagged as a biosecurity risk, and of the model declining a question as innocuous as “what does the heart do?” If your work sits anywhere near security or the life sciences — including legitimate, defensive, or educational work — expect to meet the fallback more often than the “under 5%” figure implies.

2. Silent performance throttling

This is the company's response to its own warnings about recursive self-improvement — the worry that capable models could start accelerating the development of even more capable ones. The intent is defensible; the lack of disclosure to the affected user is what drew the anger.

3. No zero-data-retention option

A notable break with Anthropic's enterprise norms: all traffic on Mythos-class models — Fable 5 included — is subject to mandatory 30-day data retention for safety monitoring, on both first- and third-party surfaces. That applies even to customers who previously held zero-retention agreements. Anthropic says the data will not be used to train models or for any non-safety purpose, that human access is logged, and that it is deleted after 30 days in almost all cases. For some regulated or privacy-sensitive organisations, the loss of a zero-retention option is a genuine blocker.

Underneath all three runs a quieter, more cultural complaint. Because the public gets safeguarded Fable while vetted partners get less-restricted Mythos 5, some users have framed the launch as a preview of “AI inequality” — the most powerful capabilities reserved for a privileged tier. It is a debate worth being aware of, even if you find the safety logic sound.

Pricing and Access

Fable 5 costs $10 per million input tokens and $50 per million output tokens — roughly double Claude Opus 4.8, and the price does not climb for longer-context usage. It also exposes graduated thinking-effort levels (from low through to “max”), and at the top setting it will happily spend large numbers of tokens reflecting on and validating its own output. That is the lever that turns a 10-cent answer into a 70-cent one.

The subscription situation is the part to flag for anyone on a paid Claude plan, because it is unusual:

• API and consumption-based Enterprise: fully available now, billed per token.
• Pro, Max, Team and seat-based Enterprise: included at no extra cost only through 22 June 2026.
• From 23 June: Fable 5 is removed from those plans and using it requires usage credits, until Anthropic restores it as a standard feature “as quickly as possible.”

In other words, if you are on a subscription, you have a roughly two-week window to evaluate Fable 5 for free before the meter starts. Worth blocking out some time deliberately.

The Bigger Picture

The timing is not accidental, and it is part of why this launch is being read as a strategic moment rather than just a product update. Anthropic confidentially filed IPO paperwork with the SEC roughly a week before the release. The company's most recent funding round valued it at around $965 billion, and it has cited a revenue run rate near $47 billion, up from roughly $10 billion a year earlier. Putting its most powerful model into the market on the eve of going public is, among other things, a statement to investors.

It also sits against an unusually candid backdrop. In the lead-up, Anthropic publicly urged AI labs to agree on coordinated mechanisms to slow frontier development if needed, warning that systems are advancing fast enough that they could soon begin improving themselves with little human involvement. Shipping the most capable public model ever while sounding that alarm is the central tension of the release — and a good reminder that the safeguards described above are not marketing dressing but the company's attempt to live with its own warnings.

Should You Switch?

For most people and teams, the sensible answer is: use Fable 5 selectively, not as your default. The case for each model breaks down cleanly by task:

• Reach for Fable 5 when the task is long, hard and agentic — large code migrations, multi-day builds, deep multi-step research — and when a wrong answer is expensive. This is where its lead is real and widens.
• Stay on Claude Opus 4.8 for everyday chat, drafting, and high-volume or latency-sensitive work. It is half the price, genuinely capable, and it is the model Fable falls back to anyway.
• Consider GPT-5.5 when cost is the deciding factor — it is roughly half Fable's price and competitive on reasoning and long-context retrieval, even where it trails on coding.
• Consider Gemini 3.1 Pro for the cheapest option and the broadest native multimodality across image, audio and video.

If you want a fuller side-by-side of the current frontier — capabilities, pricing and which to pick for what — see the companion piece, ChatGPT vs Claude vs Gemini in 2026. And if your organisation is weighing how to actually adopt and govern models like this, the Corporate Training programme covers model selection, governance and rollout planning.

Frequently Asked Questions

What is Claude Fable 5?

It is the first publicly available model from Anthropic's new Mythos tier, which sits above the Opus line. Released on 9 June 2026, it is the most capable model Anthropic has made generally available. It shares the same underlying model as Claude Mythos 5 but ships with safety classifiers that route high-risk requests to Claude Opus 4.8.

How much does Claude Fable 5 cost?

$10 per million input tokens and $50 per million output tokens via the API — about double Opus 4.8. It was free on Pro, Max, Team and seat-based Enterprise plans through 22 June 2026; from 23 June, using it on those plans requires usage credits until Anthropic restores it as standard.

Is it better than GPT-5.5 and Gemini 3.1 Pro?

On Anthropic's published benchmarks, yes — clearly on agentic coding, knowledge work, vision and legal tasks, with the lead widening on longer tasks. But those are vendor-reported numbers, GPT-5.5 costs around half as much, and Gemini 3.1 Pro is cheaper still. Cost and task fit still matter.

Why does Fable 5 sometimes fall back to Opus 4.8?

Classifiers watch for requests in three high-risk areas — offensive cybersecurity, biology and chemistry, and model distillation. When one triggers, Opus 4.8 handles the request and you are told. Anthropic says this happens in fewer than 5% of sessions, but the filters are deliberately conservative and sometimes flag benign work.

What is the difference between Fable 5 and Mythos 5?

They are the same underlying model. Fable 5 is the public version with safety classifiers in place. Mythos 5 has those safeguards lifted in certain areas and is restricted to vetted Project Glasswing partners, with a trusted-access program planned for biology researchers.

Should I switch to it?

Use Fable 5 for hard, long-running agentic work where wrong answers are expensive. For everyday chat and high-volume, latency-sensitive work, Opus 4.8 at half the price remains the sensible default.