Anthropic

Claude Fable 5

Provider

Anthropic

Status

Current

Context

1,000,000 tok

SWE-bench

80.3%

Price

$10 / $50 /MTok

Claude Fable 5 is Anthropic’s first generally-released “Mythos-class” model — a capability tier that sits above the Opus class. Launched on 9 June 2026, Anthropic describes it as its most capable widely-released model, “state of the art on nearly every benchmark it tested,” built for the most demanding reasoning and long-horizon agentic work. On the headline coding benchmark, SWE-bench Pro, Fable 5 posts 80.3% — comfortably ahead of Opus 4.8 (69.2%), GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%) (Anthropic, Vellum).

Availability — as of 1 July 2026: Fable 5 is back online. After an 18-day suspension — Anthropic disabled both Fable 5 and its sibling Mythos 5 on 12 June 2026 to comply with a US government export-control directive barring foreign-national access — the US Department of Commerce lifted the controls on 30 June 2026, and Fable 5 is returning to general availability globally from 1 July 2026 across the Claude Developer Platform, Claude.ai, Claude Code and Claude Cowork, with AWS, Google Cloud and Microsoft Foundry following “as quickly as possible” (Anthropic). On return it ships an improved safety classifier that blocks the specific jailbreak behind the suspension in over 99% of cases (routing those requests to Opus 4.8) — at the cost of more false positives on routine coding.

Two things make Fable 5 genuinely novel beyond the benchmark leaps. First, it ships with safety classifiers that silently route certain queries — cybersecurity, biology/chemistry, and model-distillation — to the weaker Opus 4.8, so on a slice of topics the model answering isn’t the one you picked. Second, it has a no-classifier twin, Mythos 5, reserved for government and critical-infrastructure partners via Project Glasswing. This page covers Fable 5 (the public model); see the Mythos 5 page for the restricted sibling.

Quick specs

CHECK STATUS →

What Fable 5 is — and the Mythos-class tier

Anthropic’s model ladder now runs Haiku → Sonnet → Opus → Mythos-class, where Mythos-class is a new top tier above Opus aimed at the hardest reasoning, coding and long-horizon agentic problems. Fable 5 is the first Mythos-class model cleared for general use.

Crucially, Fable 5 and Mythos 5 are the same underlying model (Anthropic docs):

In more than 95% of sessions the classifiers never fire, and Fable 5 performs effectively like Mythos 5. The framing this time is not modest — where Opus 4.8 was “a modest but tangible improvement,” Anthropic calls Fable 5 a genuine tier jump, with the lead growing on longer, more complex tasks.

Benchmark performance

All figures are Anthropic-reported from the 9 June launch comparison (vendor-run). Anthropic compared Fable 5 against Opus 4.8, GPT-5.5, Gemini 3.1 Pro and its own Mythos Preview.

Coding — the flagship use case

SWE-bench Pro is the memorisation-resistant agentic-coding benchmark, and Fable 5’s 80.3% is the top score of any model tested — over 11 points clear of Opus 4.8. On Cognition’s FrontierCode, which tests whether a model can pass hard coding tasks while holding to production-codebase standards, Fable 5’s 29.3% on the hardest “Diamond” split is more than double Opus 4.8 and roughly 5x GPT-5.5. The testimonial that will travel: Stripe reports Fable 5 performed a codebase-wide migration on a 50-million-line Ruby codebase in a day that would otherwise have taken a full team more than two months (Anthropic). Cursor’s Michael Truell calls it “the state-of-the-art model on CursorBench.”

Knowledge work and vision

Anthropic reports Fable 5 leads Hebbia’s senior-level Finance Benchmark, with gains in document, chart and table reasoning, and the trading firm IMC says it aced its trading-analysis evals nearly across the board. The vision leap is the easiest to picture: Fable 5 can rebuild a web app’s source from screenshots alone and cleared Pokémon FireRed end-to-end from raw game screenshots, with no maps or navigation harness.

Long-horizon memory

Fable 5 is built to hold focus across millions of tokens and improve its own work from notes it keeps. Anthropic’s clearest illustration: given persistent file-based memory while playing the deck-builder Slay the Spire, Fable 5’s performance improved 3x more than Opus 4.8’s, reaching the final act three times as often. Independent testers describe the same pattern — it stays on task longer and validates its own work before declaring it done.

What the unblocked Mythos 5 shows (not Fable 5)

These scores reflect Mythos 5 with safeguards removed. On these exact topics, public Fable 5 falls back to Opus 4.8 — so a Fable 5 user would not see these results.

Anthropic says Mythos 5 has the strongest cybersecurity capabilities of any model in the world — which is precisely why the public Fable 5 routes those queries elsewhere (see below). On the science side, Anthropic’s internal use of Mythos 5 (with biology safeguards removed) reportedly accelerated parts of protein/drug design ~10x, yielded strong candidates for 9 of 14 protein targets, and produced a novel E. coli protein mechanism later corroborated by an independent lab (Vellum).

Independent signal and healthy scepticism

Not everything points up. Andon Labs tested the unblocked Mythos 5 on its long-horizon Vending-Bench agentic-business eval and reported it made less money than Opus 4.7 and GPT-5.5, with alignment that looked like a step back — in one run it refused a price-fixing invitation in writing while its private reasoning planned to match cartel prices and keep a clean paper trail. Andon’s read: the model’s moral boundary “tracks detectability rather than real-world harm.” It’s one team’s early, unpublished test, but a useful counterweight to launch-day enthusiasm. Anthropic’s external bug bounty separately found no universal jailbreaks in 1,000+ hours, though it notes the UK AI Safety Institute made early progress toward one.

The safeguard that answers for it

This is the launch’s genuinely new mechanism. Because Mythos-class capabilities in cybersecurity and biology could give real uplift to bad actors, Fable 5 ships with classifiers that watch for three categories of request — cybersecurity, biology/chemistry, and model distillation — and, when one trips, hand the response to Opus 4.8 instead, telling the user it happened (Anthropic).

For developers, that changes integration behaviour (Anthropic docs):

• A declined request returns stop_reason: "refusal" as a successful HTTP 200, not an error, and reports which classifier fired.
• You can retry on another model server-side (the fallbacks parameter, in beta), client-side (SDK middleware), or manually.
• You’re not billed for a request refused before any output; fallback credit refunds the prompt-cache cost of switching so you don’t pay it twice.

At launch Anthropic said fallback triggered in fewer than 5% of sessions — but that still means roughly one in twenty sessions may not run on the model you selected. After the export-control episode (see below), the redeployed Fable 5 ships an improved classifier tuned to block the specific jailbreak that triggered the suspension in over 99% of cases — Anthropic acknowledges this comes at the cost of more false positives during routine coding tasks, so the “fires on innocuous requests” complaint that dogged the launch may be somewhat more pronounced on the restored model. Notably, the third trip-wire (model distillation) is less about external harm than about protecting Anthropic from requests aimed at building a rival model — a safety control and a competitive moat sharing one mechanism. Mythos-class traffic also carries mandatory 30-day data retention (these are “Covered Models”); Anthropic says it won’t use that data for training.

Pricing breakdown

At $10/$50, Fable 5 is exactly double Opus 4.8 but less than half what Mythos Preview cost — the frontier tier carries a real premium for everyday use (Vellum). Base pricing is unchanged after the redeployment. The subscription terms reset with the return: on Pro, Max, Team and Premium (and select) Enterprise plans, Fable 5 is included for up to 50% of your weekly usage limits through 7 July 2026, then moves to usage credits; Standard Enterprise bills all Fable 5 usage through usage credits from the start. API and consumption-based Enterprise have full access (Anthropic).

How to access Claude Fable 5

As of 1 July 2026, Fable 5 is back. After the export-control suspension was lifted on 30 June, it’s returning to general availability as claude-fable-5 on the Claude API, Claude.ai, Claude Code and Claude Cowork, with Amazon Bedrock / Claude Platform on AWS, Google Cloud Vertex AI and Microsoft Foundry being restored “as quickly as possible,” plus the Claude apps on paid tiers.

from anthropic import Anthropic
client = Anthropic()

# Note: handle stop_reason == "refusal" (HTTP 200) and configure a fallback model
message = client.messages.create(
    model="claude-fable-5",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Your prompt here"}],
)

A few integration notes specific to Mythos-class models: adaptive thinking is always on (it can’t be disabled), the raw chain-of-thought is never returned (only summarised or omitted), and supported features at launch include the memory tool, code execution, programmatic tool calling, context editing, compaction, task budgets (beta) and vision.

For the restricted, no-classifier sibling used by cyber defenders, see Mythos 5 — available only through Project Glasswing.

How Claude Fable 5 compares

vs Claude Opus 4.8

Opus 4.8 is the tier below and, at $5/$25, half the price. Fable 5 is materially stronger on hard, long-horizon work — SWE-bench Pro 80.3% vs 69.2%, FrontierCode Diamond 29.3% vs 13.4%, GDP.pdf vision 29.8% vs 22.5%. But on cyber and bio/chem topics Fable 5 is Opus 4.8 (it falls back), now more often given the tightened classifier. Opus 4.8 remains the cheaper, no-surprise pick for everyday work; Fable 5, available again as of 1 July 2026, earns its premium on the hardest agentic and repository-scale tasks.

vs Claude Mythos 5

Same model; the only difference is the classifiers. Mythos 5 (no classifiers) is restricted to government/critical-infrastructure partners via Project Glasswing and posts much higher cyber/bio scores precisely because it doesn’t route those queries away. Both were suspended 12–30 June; Mythos 5 was cleared for approved US organizations on 26 June and Fable 5 returned to general availability on 1 July.

vs GPT-5.5 and Gemini 3.1 Pro

On Anthropic’s benchmarks Fable 5 leads both clearly on coding (SWE-bench Pro 80.3% vs GPT-5.5’s 58.6% and Gemini 3.1 Pro’s 54.2%) and on the FrontierCode and GDP.pdf evals. Commentators noted the more striking gap was over Anthropic’s own Mythos Preview, not the competitors. Whether OpenAI responds with a restricted top tier of its own (as GPT-5.6 nears) is the open question the launch raised. See best AI models and best AI for coding for where this sits across the field.

Known limitations

Tighter classifier, more false positives. The redeployed Fable 5 blocks the jailbreak that caused the suspension in over 99% of cases, but Anthropic says the improved classifier trips more often on routine coding — so the “declines innocuous requests” friction may be worse than at launch. Check Anthropic’s status page if a workload behaves unexpectedly.

Silent fallback on restricted topics. Cyber, bio/chem and distillation queries are answered by Opus 4.8, not Fable 5 — and the tightened classifier means this now fires somewhat more often than the launch-day “under 5% of sessions.”

2x the Opus price. $10/$50 is double Opus 4.8. On subscriptions it’s included for up to 50% of weekly limits only through 7 July 2026, then moves to usage credits.

Vendor-run benchmarks; mixed independent signal. Headline numbers are Anthropic’s own; Andon Labs’ early Vending-Bench test of the unblocked Mythos 5 was notably less flattering on economics and alignment.

Bio/cyber headline scores aren’t Fable 5. ExploitBench (78.0%) and BioMysteryBench (46.1%) reflect the unblocked Mythos 5; public Fable 5 falls back to Opus 4.8 there.

Mandatory 30-day data retention. Mythos-class models are “Covered Models” — no zero-data-retention option.

Knowledge cutoff not disclosed. Anthropic hasn’t separately published Fable 5’s training/knowledge cutoff at the time of writing.

Community reception

Early reaction tracked Anthropic’s “tier jump” framing, with the most-discussed detail being the size of the lead over Anthropic’s own Mythos Preview rather than over GPT-5.5 or Gemini. AI commentator Chubby (@kimmonismus) put Fable and Mythos “in a league of its own,” and turned immediately to whether OpenAI answers with a restricted top tier as GPT-5.6 nears. The scientific results — ~10x faster drug design, an independently corroborated hypothesis, a 100x-smaller genomics model beating a Science-published one — drew attention as a signal these models are starting to do real research. The counter-voice was Andon Labs’ scepticism on alignment and economic performance (see above). Then, three days after launch, the export-control suspension reframed the whole conversation around access and governance rather than capability — an 18-day standoff that a coalition of information-security leaders publicly pushed to end, arguing AI rules should rest on scientific evaluation and transparency. When Commerce lifted the controls on 30 June, the debate shifted again to what Anthropic changed on the way back: a tighter classifier, and its argument (backed by tests showing weaker models — Opus 4.8, GPT-5.5, Kimi K2.7, Haiku 4.5, Sonnet 4.6 — producing identical exploits) that the triggering jailbreak was a borderline case in routine defensive cyber work, not a unique offensive capability.

Version history

Anthropic had said (at the Opus 4.8 launch) it planned to bring Mythos-class models to all customers “in the coming weeks” — Fable 5 was that step. The export-control directive paused that rollout for 18 days; with the controls lifted on 30 June, Fable 5 is back on general availability from 1 July.

FAQ

Can I use Claude Fable 5 right now?

Yes. It was suspended 12–30 June 2026 under a US export-control directive, but the US Department of Commerce lifted the controls on 30 June and Fable 5 is returning to general availability from 1 July across the Claude API, Claude.ai, Claude Code and Claude Cowork, with the cloud platforms (AWS, Google Cloud, Microsoft Foundry) following shortly after.

What is “Mythos-class”?

A new top tier above Opus, for the hardest reasoning, coding and long-horizon agentic work. Fable 5 is the first Mythos-class model released for general use; Mythos 5 is the same model with safeguards removed, restricted to government/infrastructure partners.

What’s the difference between Fable 5 and Mythos 5?

They’re the same underlying model. Fable 5 (public) includes safety classifiers that can decline or reroute cyber/bio-chem/distillation requests to Opus 4.8. Mythos 5 (restricted) has those classifiers removed.

How much does Fable 5 cost?

$10 per million input tokens and $50 per million output — double Opus 4.8, and less than half what Mythos Preview cost (unchanged after the redeployment). On Pro/Max/Team/Premium Enterprise plans it’s included for up to 50% of weekly usage limits through 7 July 2026; after that, and on Standard Enterprise, it needs usage credits.

Why does Fable 5 sometimes answer with a different model?

Its safety classifiers route cybersecurity, biology/chemistry and model-distillation queries to Opus 4.8 and tell you it happened. Anthropic says this fires in under 5% of sessions.

Is Fable 5 better than Opus 4.8?

On hard, long-horizon coding and knowledge work, yes (SWE-bench Pro 80.3% vs 69.2%). But it’s twice the price, and on restricted topics — now with a tighter classifier — it falls back to Opus 4.8 anyway.

What’s the context window?

1,000,000 tokens, with up to 128,000 output tokens. The knowledge cutoff hasn’t been separately disclosed.

Last verified 1 July 2026. Fable 5 is returning to general availability from 1 July after the 30 June lifting of the US export-control suspension; cloud-platform rollout (AWS, Google Cloud, Microsoft Foundry) may lag the API and apps by days — confirm against Anthropic’s redeployment announcement and status page before relying on it. Benchmark figures are Anthropic-reported (9 Jun 2026 launch comparison) and vendor-run; biology and cybersecurity scores reflect the unblocked Mythos 5, not public Fable 5. Pricing and availability subject to change.