Claude Fable 5 Cheatsheet

Claude Fable 5 is Anthropic's first publicly available Mythos-class model, launched Tuesday, June 9, 2026. Mythos-class is a new capability tier that sits above the Opus class, and Fable 5 is the same underlying model as the restricted Claude Mythos 5, wrapped in safeguards that make it safe for general use.

In plain terms: this is the most capable model Anthropic has ever made broadly available. It is state-of-the-art on nearly every benchmark it was tested on, and the gap over Opus 4.8 widens as tasks get longer and more complex.

There is a catch, and it is the defining feature of the release. On a narrow set of high-risk topics (cybersecurity, biology and chemistry, distillation), Fable 5 will not answer with its full Mythos-class capability. It routes those queries to Opus 4.8 instead. So the model you call is genuinely a tier above Opus most of the time, and effectively Opus 4.8 on the topics it guards.

Quick Verdict

Use Fable 5 when the task is long, complex, and the quality of the result is worth paying double the Opus rate:

• codebase-scale migrations and refactors that would take a team weeks
• long-horizon agent runs where the model has to stay focused across millions of tokens
• vision-heavy work: reading dense figures, rebuilding a UI from screenshots
• senior-level knowledge work (finance, legal redlines, analytics) where accuracy pays for itself
• prototyping and "one-shot a full app" workflows where fewer turns saves real money

Stay on Opus 4.8 or Sonnet for everyday coding, cheap bulk automation, and anything in the guarded domains, where Fable 5 falls back to Opus 4.8 anyway. The price is $10/$50 per million tokens, so reach for Fable 5 when the task earns it, not as your default.

Key Specs

What a Mythos-Class Model Actually Is

Anthropic ships model families in tiers. Haiku, Sonnet, and Opus is the ladder you know. Mythos is a new rung above Opus, and it is the tier Anthropic has been running internally and with vetted partners since April.

Mythos started as Claude Mythos Preview, released only to a limited group of cyber defenders and critical infrastructure providers under Project Glasswing. Anthropic said at the time it did not plan to make Mythos generally available, because a model this strong at finding and exploiting software vulnerabilities is dangerous in the wrong hands.

Fable 5 is Anthropic following through on its stated goal of releasing Mythos-level capability to everyone, now that the safeguards are strong enough. The way it threads that needle is the whole story of this launch: one set of weights, two names.

Claude Mythos 5 is the raw model with safeguards lifted, still restricted to trusted partners. Claude Fable 5 is the exact same model with production safeguards on top, available to anyone. When a request is clean, you get the full Mythos-class model. When it touches a guarded topic, the safeguards catch it and the answer comes from Opus 4.8 instead.

The Headline Change: A Real Tier Above Opus

Most point releases lead with a few percentage points. Fable 5 leads with a tier jump.

The clearest number is SWE-bench Verified, where Fable 5 posts 95.0% against Opus 4.8's 88.6%. That benchmark is close to saturated, so the more meaningful coding signal is SWE-bench Pro, the harder agentic-coding set, where Fable 5 jumps to 80.0% from Opus 4.8's 69.2%. That is a 10.8-point gain on the benchmark that best tracks real-world agent work.

The pattern repeats on the hard, less-saturated evals. On Cognition's FrontierCode (the Diamond split), Fable 5 hits 29.3%, more than double Opus 4.8's 13.4% and far ahead of GPT-5.5's 5.7%. Anthropic reports Fable 5 scores highest among frontier models on FrontierCode even at medium effort.

The thing to internalize is the shape of the lead, not any single score. Anthropic's own framing is that the longer and more complex the task, the larger Fable 5's lead. On a quick edit, the gap over Opus 4.8 is small. On a multi-hour, multi-file, codebase-scale job, the gap is the reason to pay for it.

Why "Longer Tasks, Bigger Lead" Is the Right Headline

This release is built for autonomy, and the benchmark spread says so.

Stripe ran an early test that is the cleanest illustration. On a 50-million-line Ruby codebase, Fable 5 performed a codebase-wide migration in a day that would otherwise have taken a full team over two months by hand. That is not a model writing a function faster. That is a model holding a plan together across an enormous codebase without losing the thread.

Two capabilities make that possible. The first is sustained focus across millions of tokens. The second is memory: Fable 5 improves its own work using notes it keeps along the way. Anthropic's clearest demo is a game. Given persistent file-based memory while playing the deck-builder Slay the Spire, Fable 5's performance improved three times more than Opus 4.8's, and it reached the final act three times as often. The memory is what turns a long run into a compounding one instead of a drifting one.

Customers describe the same thing in production. Rakuten put it directly: "At the highest effort, Fable reflects on and validates its own work. For us, that's what makes highly autonomous operations possible. The extra thinking pays for itself." That self-validation is what lets you hand off a long job and trust the result.

Vision Took the Biggest Leap

Fable 5 is the new state-of-the-art for vision, and the examples are the easiest part of the launch to picture.

It can extract precise numbers from detailed scientific figures, and it can rebuild a web app's source code from screenshots alone. On GDP.pdf, a vision eval that asks the model to reason over a rendered document with no tools, Fable 5 leads at 29.8%, ahead of GPT-5.5's 24.9% and Opus 4.8's 22.5%.

The headline demo is Pokemon. Earlier Claude models needed a complex helper harness (maps, navigation aids, game-state info) to play Pokemon at all. Fable 5 beat Pokemon FireRed start to finish with a minimal, vision-only harness, working from raw game screenshots. Less scaffolding, more capability. If you build agents that read screens, this is the upgrade that matters most.

Fable 5 vs Opus 4.8

The story is "a tier up on hard, long work, at double the price," not "a little better at everything."

All benchmark figures are self-reported by Anthropic. Read the spread this way: Fable 5 leads everywhere the safeguards stay quiet, and the lead is widest on the hardest, longest tasks. On the guarded topics, you are effectively running Opus 4.8, so there is no reason to pay the Fable premium for that work.

The Safeguards Are the Defining Feature

This is the part with no equivalent in any Opus release.

Fable 5 runs separate classifier systems over every request. When a classifier fires, the main model does not answer. Three categories are in scope: cybersecurity, biology and chemistry, and distillation (attempts to extract Claude's capabilities to train competing models).

In the Claude apps, a flagged request transparently falls back to Opus 4.8 and you are told it happened. Anthropic's reasoning is that an Opus 4.8 answer is a far better experience than an outright refusal. The safeguards are tuned conservatively, so they sometimes catch harmless requests, and Anthropic says it will narrow the false positives after launch.

The frequency matters. The classifiers trigger in less than 5% of sessions on average, and more than 95% of Fable sessions involve no fallback at all. For that 95%-plus, Fable 5's performance is effectively identical to Mythos 5. So for normal product and coding work, you are getting the full Mythos-class model nearly all the time.

One subtle trap shows up in benchmark tables. When you see Fable 5 quoted at a high number on a cybersecurity or biology benchmark, that is really the Mythos 5 score. The public Fable 5 you call will trip its safeguards on those tasks and perform closer to Opus 4.8. On Terminal-Bench, for example, about 21% of trials hit a safety refusal and reverted to Opus 4.8 for the rest of the run.

The One Safeguard You Cannot See

There is a fourth intervention, and it is the one that generated the loudest reaction.

Separate from the three transparent classifiers, Anthropic added safeguards that limit Fable 5's effectiveness on requests targeting frontier LLM development: building pretraining pipelines, distributed training infrastructure, or ML accelerator design. Using Claude to build a competing model already violates the Terms of Service, and this enforces it.

The difference is that these interventions are not visible to you. Fable 5 does not fall back to another model and does not tell you anything changed. Instead it quietly limits its own effectiveness through prompt modification, steering vectors, or parameter-efficient fine-tuning. Anthropic estimates this affects roughly 0.03% of traffic, concentrated in fewer than 0.1% of organizations.

Almost no one will hit it. But it is worth knowing it exists, because it is the first time Anthropic has shipped a safeguard that silently degrades answers rather than refusing or routing them. If your work is anywhere near ML systems research, factor it in.

How the API Changed

If you are calling Fable 5 from the Messages API, several knobs you are used to are gone, and one default flipped.

• Adaptive thinking only. Thinking is always on. The model decides how much to think per request via thinking: {type: "adaptive"}. You control depth with effort levels (low through max), not a token budget.
• You cannot disable thinking. Sending thinking: {type: "disabled"} returns a 400 error. This is unique to Fable 5, so omit the parameter entirely rather than setting it to disabled.
• budget_tokens is removed. Use effort instead.
• temperature, top_p, and top_k are removed. The sampling controls are no longer accepted.
• Assistant prefills return a 400. You cannot prefill the assistant turn.
• API safeguards default to block. On the Messages API, a flagged request is refused unless you opt into the Opus 4.8 fallback yourself, server-side or via SDK middleware.

The migration takeaway: if you are porting a request from Opus 4.8, strip the sampling params and budget_tokens, drop any assistant prefill, and decide upfront how you want to handle the safeguard fallback in your own stack.

Pricing and the Usage Trap

Fable 5 is $10 per million input tokens and $50 per million output tokens.

So Fable 5 is double Opus 4.8, and less than half what the restricted Mythos Preview cost. That framing makes it sound cheap for Mythos-class capability, and per token it is.

The trap is not the rate, it is the token volume. Fable 5 is a reasoning-heavy model that thinks longer and generates far more tokens per request. Early subscription users reported burning through usage limits fast (one Reddit thread described roughly 2% of a plan per minute on heavy runs). Anthropic had to reset rate limits across products after demand spiked on launch day. Measure your actual spend per task, not the headline rate, and remember the model is most cost-efficient when its longer thinking saves you turns rather than padding short ones.

Availability and the June 22 Cliff

Where you can use Fable 5 depends on how you pay.

On the Claude API and consumption-based Enterprise plans, Fable 5 is fully available from day one. Subscriptions are rolling out in stages because demand is high and hard to predict:

• Through June 22: included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost.
• June 23: Fable 5 is removed from those plans. Using it after that requires usage credits. Anthropic may extend the included window if capacity allows.
• Later: when capacity is sufficient, Anthropic intends to restore Fable 5 as a standard part of subscription plans, as quickly as it can.

If you want to evaluate Fable 5 on the cheap, the window before June 23 is the time to do it. After that, subscription access is metered until capacity catches up.

Claude Mythos 5, the unsafeguarded sibling, stays restricted to Project Glasswing partners (cyber safeguards lifted) and, soon, select biology researchers (bio and chem safeguards lifted) through a trusted access program.

Should You Use Claude Fable 5?

Yes, if your work is:

• codebase-scale migrations and refactors, the Stripe-shaped jobs that take teams weeks
• long-horizon agent runs where staying focused across millions of tokens is the whole game
• vision-heavy tasks: dense figures, screenshots-to-code, screen-reading agents
• senior-level knowledge work where a sharper answer is worth double the rate
• prototyping where one-shotting an app beats a hundred prompts

Probably not your default if your workload is everyday edits, cheap bulk automation, or anything in the guarded domains, where Fable 5 falls back to Opus 4.8 and the premium buys you nothing. For most teams the move is the familiar one: keep Sonnet and Opus 4.8 for daily work, and reach for Fable 5 on the long, hard, high-stakes jobs where its lead is widest. That is exactly where this release was built to win.

Frequently Asked Questions

What is Claude Fable 5?

Claude Fable 5 is Anthropic's first publicly available Mythos-class model, released June 9, 2026. Mythos-class is a new capability tier above the Opus class. Fable 5 is the same underlying model as the restricted Claude Mythos 5, with production safeguards added so it is safe for general use.

What is a Mythos-class model?

Mythos is a tier above Opus in Anthropic's model lineup. It first appeared as Claude Mythos Preview in April 2026, restricted to cyber defenders and infrastructure providers under Project Glasswing. Fable 5 and Mythos 5 are the same weights: Mythos 5 has safeguards lifted and stays restricted, Fable 5 carries the safeguards and is generally available.

Is Claude Fable 5 free?

It is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans through June 22, 2026. On June 23 it is removed from those plans and using it requires usage credits, with Anthropic aiming to restore it to subscriptions once capacity allows. On the Claude API and consumption-based Enterprise plans it is available from day one at $10 input and $50 output per 1M tokens.

When was Claude Fable 5 released?

Tuesday, June 9, 2026. It launched alongside Claude Mythos 5, the restricted sibling model with safeguards lifted for trusted partners.

Why does Claude Fable 5 sometimes answer with Opus 4.8?

Fable 5 runs classifiers for cybersecurity, biology and chemistry, and distillation. When one fires, the request is routed to Opus 4.8 instead of the full Mythos-class model, and in the Claude apps you are told it happened. This triggers in less than 5% of sessions, so more than 95% of the time you get Fable 5's full capability.

How much better is Claude Fable 5 than Opus 4.8?

A tier better on long, hard tasks, and roughly even on short ones. SWE-bench Pro jumps from 69.2% to 80.0%, and FrontierCode Diamond more than doubles from 13.4% to 29.3%. Anthropic's framing is that the longer and more complex the task, the larger Fable 5's lead, and it costs twice as much per token.

Sources

• Claude Fable 5 and Claude Mythos 5 (Anthropic)
• Anthropic's Claude Fable 5 is a version of Mythos the public can access today (TechCrunch)
• Anthropic releases Mythos-like AI model to the public (CNBC)
• Anthropic Launches Claude Fable 5, Its First Public Mythos-Class Model (MacRumors)
• Claude Fable 5: Review, Benchmarks and Pricing (LLM Stats)
• Claude Fable 5 & Mythos 5 Benchmarks Explained (Vellum)
• If Claude Fable stops helping you, you'll never know (Simon Willison)

Related Pages

• Claude Opus 4.8
• Claude Mythos and OpenMythos
• Claude Code Models
• Claude Code Pricing and Token Usage