Claude Fable 5 Explained — Anthropic Goes Beyond Opus

Originally published June 9, 2026, the day Fable 5 launched. Updated since — including a full account of the June 12 suspension and July 1 restoration in the Update section below.

Claude Fable 5 is the first model of Anthropic’s Claude 5 family and the most intelligent Claude you can actually get your hands on. It is a “Mythos-class” model — a capability tier that sits above Opus — made safe enough for general availability, priced at $10 per million input tokens and $50 per million output tokens with a 1M-token context window. I have had it in my daily rotation since launch day (minus the 19 days it went dark — more on that below), and this guide covers everything: what it is, how it behaves in the API, how it differs from its restricted sibling Mythos 5, how to access it on every surface, and when it is genuinely worth paying double the price of Opus 4.8.

TL;DR:

• Claude Fable 5 launched June 9, 2026 as the first Claude 5 model — a Mythos-class model above Opus, and Anthropic’s most intelligent generally available model.

• API: claude-fable-5, $10/$50 per MTok, 1M context, 128K max output, thinking always on, no temperature/top_p/top_k.

• It refuses more than Opus: safety classifiers cover offensive cyber, dual-use bio/chem, and distillation attempts, triggering in under 5% of sessions.

• It was suspended globally from June 12 to July 1, 2026 under US export controls, then restored with a stricter classifier.

• For most work, Opus 4.8 at half the price is still the right default. Fable 5 earns its premium on long-horizon, genuinely hard tasks.

What Is Claude Fable 5?

Claude Fable 5 is Anthropic’s flagship model, announced on June 9, 2026 alongside its restricted sibling, Claude Mythos 5. Two things make it different from every Claude release before it.

The first model of the Claude 5 family

Fable 5 is the first Claude 5 generation model to ship. Notably, Anthropic did not lead the new generation with an Opus — it led with something above Opus. The familiar Opus, Sonnet, and Haiku tiers continue (Sonnet 5 followed on June 30), but Fable establishes a new top rung on the ladder.

A Mythos-class model made safe for general availability

Anthropic describes Mythos-class models as “a tier of Claude models that sit above our Opus class in capability.” The class debuted quietly in April 2026 as Claude Mythos Preview, available only to vetted partners. Fable 5 is the same class of model wrapped in additional safeguards so that anyone with a Claude account or API key can use it. That wrapping is not cosmetic — it changes refusal behavior in specific domains, which I cover in the API section below.

In practice, “above Opus” means Fable 5 outperforms Opus 4.8 on nearly all of Anthropic’s tested benchmarks, with the gap widest on complex, extended tasks. Anthropic reports it scored highest among frontier models on the FrontierCode evaluation, completed Pokémon FireRed through a vision-only interface (prior models needed helper scaffolding), and that persistent memory improved its results on the long-horizon game benchmark Slay the Spire about three times more than it improved Opus 4.8’s, with Fable reaching the final act three times as often.

Update: The Suspension and Restoration (June 12 – July 1, 2026)

Three days after launch, Fable 5 disappeared. On June 12, 2026, the US government applied export controls to Fable 5 and Mythos 5, reportedly after an Amazon researcher’s report described a jailbreak technique that could elicit software-vulnerability identification from the model. The order restricted access for foreign nationals effective immediately — and because Anthropic had no way to verify any user’s nationality in real time, it suspended both models for everyone, worldwide.

The controls were lifted on June 30, and on July 1 Anthropic restored Fable 5 globally — on the Claude Platform, claude.ai, Claude Code, and Claude Cowork — with a new safety classifier it says blocks the reported technique in more than 99% of cases. In my use since restoration, the model itself feels unchanged; the only visible difference is the occasional false-positive refusal on security-adjacent coding work. I wrote up the full incident, day by day, in the dedicated Claude Fable 5 suspension retrospective — worth reading if you build products on the API, because the lessons about single-model dependency apply to every provider.

Claude Fable 5 Capabilities: What It Is Actually Better At

Benchmark tables are one thing; a month of real use (interrupted, admittedly) is another. Here is where I have seen Fable 5 clearly separate from Opus 4.8:

• Long-horizon agentic work. Tasks that run across hours and millions of tokens — large refactors, multi-repo migrations, extended research syntheses — are where the gap is unmistakable. Fable 5 keeps a coherent plan alive far longer before drifting. Anthropic cites Stripe compressing a codebase migration estimated at two months of team effort into a single day.
• Self-checking at high effort. At the highest effort settings, Fable 5 reflects on and validates its own work before presenting it. In my testing this shows up as fewer “confidently wrong” answers on hard problems, at the cost of longer waits.
• Scientific and knowledge work. Anthropic reports state-of-the-art results in software engineering, knowledge work, vision, and scientific research, including a roughly 10x acceleration in protein-design tasks during pre-release partner testing.
• Vision-driven autonomy. The Pokémon FireRed result matters less for the game and more for what it implies: the model can operate a UI from raw screenshots for extended periods without hand-built helpers.

What it is not dramatically better at: short chats, routine coding questions, summarization, everyday writing. On a 40-line bug fix or a five-paragraph email, I genuinely cannot tell Fable 5 and Opus 4.8 apart often enough to justify the price difference.

Claude Fable 5 in the API: Pricing, Thinking, and Refusals

The API model ID is claude-fable-5. Here is how it lines up against its neighbors (see the full Claude API pricing breakdown for every tier and caching multiplier):

Model	API ID	Input / output per MTok	Context	Max output
Claude Fable 5	claude-fable-5	$10 / $50	1M	128K
Claude Opus 4.8	claude-opus-4-8	$5 / $25	1M	128K
Claude Sonnet 5	claude-sonnet-5	$3 / $15 (intro $2 / $10 through Aug 31, 2026)	1M	128K
Claude Haiku 4.5	claude-haiku-4-5	$1 / $5	200K	64K

Thinking is always on

Fable 5 thinks on every request — there is no way to disable reasoning. You steer how hard it thinks with the effort parameter (low through max) rather than a token budget; fixed budget_tokens is gone on current models. Like Opus 4.7+ and Sonnet 5, Fable 5 also drops the classic sampling knobs: temperature, top_p, and top_k are removed, and requests that include them error. If your integration still sets temperature, fix that before switching models.

Refusal handling in dual-use domains

This is the part most writeups skip, and it is the most important operational difference from Opus. Because Fable 5 is a Mythos-class model, Anthropic runs additional safety classifiers on it in three areas:

• Cybersecurity — offensive-leaning tasks (exploit development, vulnerability discovery framing) get blocked.
• Biology and chemistry — conservative blocking on dual-use queries.
• Distillation — attempts to extract the model’s capabilities at scale get flagged.

When a classifier fires, the request can fall back to Opus 4.8 rather than failing outright. Anthropic says the safeguards trigger in under 5% of sessions, and that external red-teaming found no universal jailbreaks in over 1,000 hours of testing. My experience matches the under-5% figure — but if you work in security, expect the false positives to cluster in exactly your domain, and slightly more so since the stricter post-restoration classifier shipped on July 1. Build your error handling to detect the fallback rather than assuming every response came from Fable.

Fable 5 vs Mythos 5 and Project Glasswing

Claude Mythos 5 is the same underlying model as Fable 5 with Fable’s additional dual-use safety measures removed. It is not something you can sign up for: Mythos 5 deploys through Project Glasswing, Anthropic’s collaboration with the US government focused on defensive cybersecurity, which began with Claude Mythos Preview in April 2026. Approved organizations only, with a broader trusted-access program planned later.

Both models cost the same $10/$50 per MTok — less than half what Mythos Preview cost — which tells you the price premium over Opus reflects the capability tier, not the safety wrapper. For everyone outside Glasswing, “Fable 5 vs Mythos 5” is a non-decision: Fable is the version that exists for you. The distinction matters mainly for understanding the June suspension, since the export-control order named both models.

How to Access Claude Fable 5 on Every Surface

claude.ai, desktop, and mobile

On paid plans (Pro, Max, Team, Enterprise), Fable 5 appears in the model picker like any other model. It is not available on the Free plan. Extended Thinking controls do not apply the way they do on other models — Fable thinks regardless.

Claude Code and Claude Cowork

In Claude Code, switch with /model claude-fable-5 (or pick it in the desktop/web UI). It is also selectable in Claude Cowork, including the Windows research preview. Note that fast mode (/fast) is an Opus feature — it does not apply to Fable 5.

Claude API

Request claude-fable-5 via POST https://api.anthropic.com/v1/messages with your key from platform.claude.com. Cloud-provider availability (AWS, Google Cloud, Microsoft Foundry) has been trailing first-party access since the July 1 restoration.

Plan-inclusion mechanics: read this before assuming it is “included”

Fable 5’s subscription access has been window-based rather than permanent, and the windows have moved twice already:

• June 9–22 (as launched): included at no extra cost on paid tiers, with a staged rollout; usage credits were to be required after June 23 until capacity caught up. The suspension interrupted this window.
• July 1–7 (post-restoration): Pro, Max, Team, and select Enterprise plans got Fable 5 for up to 50% of their weekly usage limits at no extra cost.
• After July 7: Fable 5 usage on subscriptions draws on usage credits. API access is straightforward pay-per-token throughout.

If you are weighing a subscription primarily for Fable access, check the current mechanics in-app first — this is the fastest-moving detail in this article, and I expect it to change again as capacity grows.

Is Claude Fable 5 Worth It vs Opus 4.8?

Honest answer: for most people, most of the time, no — Opus 4.8 at $5/$25 remains the right default for demanding work. Here is the decision logic I actually use:

• Choose Fable 5 when the task is long-horizon and failure is expensive: multi-hour agent runs, large migrations, hard research problems, work where a wrong answer costs you more than double tokens. One task that succeeds in a single Fable run instead of three Opus retries is cheaper in both money and time.
• Choose Opus 4.8 for everyday demanding work: solid coding, strong writing, half the price, fast mode support, and no dual-use classifier layer. It is also the model Fable falls back to anyway when a classifier fires.
• Choose Sonnet 5 or Haiku 4.5 for volume work — near-Opus coding quality at $2/$10 intro pricing makes Sonnet 5 the value play right now.

There is also an availability argument the suspension taught everyone: Fable 5 is the newest, most scrutinized model in the lineup, and it has already spent 19 days offline. If your product cannot tolerate that, either default to Opus or build Fable in with an automatic fallback.

What People Get Wrong About Fable 5

• “Fable 5 is the new Opus.” No — it is a tier above Opus. The Opus line continues, and Opus 4.8 remains Anthropic’s recommendation for most demanding work.
• “Mythos 5 is a bigger model.” It is the same model as Fable 5 with fewer safeguards, restricted to approved organizations. Capability differences show up only where Fable’s classifiers intervene.
• “You can turn thinking off to make it cheaper.” You cannot. Thinking is always on; you control depth with effort, not a switch.
• “The suspension was Anthropic pulling a broken model.” It was a US export-control order, and the global scope came from the impossibility of real-time nationality verification — not from a model defect. The full story is in the suspension retrospective.
• “$10/$50 makes every task expensive.” Prompt caching still applies (cache reads at roughly a tenth of input price), and always-on thinking is token-efficient compared to earlier reasoning modes. Measure before assuming.

FAQ

Where to Go Next

Fable 5 is the most capable Claude yet, but it is one model in a fast-moving lineup — a lot has shipped this year, from Sonnet 5’s intro pricing to fast mode on Opus. For the full picture of everything Anthropic has released in 2026 and how the pieces fit together, start with our regularly updated Claude AI news roundup, and browse the rest of our Claude guides for the fundamentals.