Claude Fable 5

Spec

Attribute	Value
Developer	Anthropic
Tier	Mythos-class (first publicly available)
Released	2026-06-09
Announced	2026-06-09
Context window	1M tokens
Pricing	$10/M input · $50/M output
License	proprietary (API-only; no weight release)
Safety gating	Yes — <5% of sessions route sensitive queries to Claude Opus 4.8
Availability	Claude.ai, Claude API, AWS Bedrock, GitHub Copilot, Google Cloud Vertex AI
Free period	Pro / Max / Team / seat Enterprise: free through 2026-06-22 (2× usage credits after)

Release Date

June 9, 2026 — the first publicly available Mythos-class model from Anthropic. Fable 5 and Claude Mythos 5 share one underlying model. Fable 5 wraps the model in classifier gates for general availability; Mythos 5 operates with some gates lifted exclusively for approved Project Glasswing partners (cyberdefenders, infrastructure providers, US government).

Status — restored globally (2026-07-01)

Restoration (2026-06-30 / 2026-07-01)

The US Department of Commerce lifted export controls on June 30, 2026 — 19 days after imposing them. Fable 5 was restored to global users on Claude.ai, Claude Code, Claude Cowork, and the API starting July 1, 2026. Conditions of restoration:

• Fable 5 capped at up to 50% of weekly usage limits through July 7; Anthropic extended free access through July 12, then again through July 19, 2026 at 11:59 PM PT (third extension total, announced July 13 just as the July 12 deadline hit). Terms of the July 19 extension: Pro/Max/Team/Enterprise, no extra cost, +50% rate limit boost. After July 19, Fable 5 credits-only at $10/$50 per Mtok. Anthropic states the subscription restriction is temporary pending capacity expansion. (source)
• Anthropic agreed to proactively detect and address security risks, help develop frontier-model standards, and report any malicious activity to the US government
• AWS, Google Cloud, Microsoft Foundry re-enabling access as quickly as possible

Mythos 5 is being restored to a broader set of US organizations (beyond the ~100–150 cleared on June 26). → Anthropic (source) (CNBC) (Forbes)

Suspension (2026-06-12 to 2026-07-01, 19 days)

Three days after launch, on 2026-06-12 the US government issued an export-control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 for any foreign national worldwide (including Anthropic's own foreign-national employees). Anthropic disabled both models for all customers to comply. The government cited a claimed jailbreak method (Amazon researchers demonstrating vulnerability identification and exploit-code generation); Anthropic reviewed the demonstration and found only minor, previously-known vulnerabilities, disputing that this warrants recalling a model deployed to hundreds of millions. Reported as the first US export ban on a named AI model. → (source) (statement)

Benchmarks

Benchmark	Fable 5	Opus 4.8	GPT-5.5	Gemini 3.1 Pro
SWE-bench Pro	80.3%	69.2%	58.6%	54.2%
SWE-bench Verified	95.0%	88.6%	—	—
Terminal-Bench 2.1	88.0%	82.7%	83.4%	70.7%
GDPval-AA (reasoning)	1932	1890	1769	1314
FrontierCode Diamond	29.3%	13.4%	5.7%	—
Notable: the FrontierCode Diamond gap over Opus 4.8 (+15.9 pp) and over GPT-5.5 (+23.6 pp) is unusually large — this benchmark targets the hardest real-world unsolved coding problems.

Fable 5 also surpasses Mythos Preview on SWE-bench Pro (80.3% vs 77.8%), suggesting post-training improvements were made between the April preview and the June GA.

(sources), (BenchLM), (MangoMind)

Use Cases

• Long-horizon software engineering — Stripe case study: codebase-wide migration completed in 1 day vs. 2+ months for a full team
• Complex knowledge work: coding, science, vision, research
• Agentic workflows running for longer than any previous Claude model
• General-purpose reasoning on the hardest tasks (FrontierCode Diamond leader)

Safety Architecture

A two-model approach:

1. Default path: Fable 5 handles the query directly
2. Safety gate (< 5% of sessions): sensitive queries are silently routed to Claude Opus 4.8

This design delivers Mythos-class capability for ≥95% of use cases while preventing misuse in the remaining edge cases. It represents a middle path between "fully open" (Mythos Preview would be unsafe) and "fully restricted" (Mythos Preview was withheld entirely).

Cyber Safeguards (2026-07-01, added at restoration)

At the time of global restoration (July 1, 2026), Anthropic deployed a new safety classifier blocking >99% of attempts replicating the specific jailbreak that triggered the June 12 export-control suspension. Broader behavioral guardrails were also extended.

Biology Safeguards Retrained (2026-08-07)

Anthropic published an update to Fable 5's biology safeguards that cuts biology-related fallbacks by about 85% in testing across its product surfaces (source).

The mechanism changed is the classifier's constitution — the rule set deciding what counts as safeguarded content. Anthropic rewrote and retrained it with detailed exceptions for benign use cases, feedback from internal and external experts, and new training data reflecting the revised rules. The stated aim is to separate everyday health and educational questions — interpreting lab results, understanding symptoms, learning biology — from dual-use research.

Expected reduction in total fallback volume, by surface:

Surface	Expected reduction
Claude.ai	~67%
Cowork	~55%
Claude Code	~17%
Claude Platform	~7%
Two things this does not change. Dual-use work — **virology, toxicology and
molecular design** — still falls back. And the fallback target named in this post is
Claude Opus 5, where the model's original safety gate
(above, and in the Safety gating spec row) routed to Claude Opus 4.8; no source
read here states when that target changed, so both are recorded rather than
reconciled (source).

The spread across surfaces is the part worth keeping: an intervention that removes two-thirds of fallbacks on Claude.ai and one-fourteenth on the Platform is a statement about who was being falsely blocked. The classifier was firing on consumer questions, not on developer traffic.

Mandatory 30-day data retention (2026-06-09,

From launch day, all traffic to Fable 5 — and to Claude Mythos 5 — is retained for 30 days for trust and safety, on both first- and third-party surfaces. Fable 5 does not support Zero Data Retention, unlike other Claude API models, and the requirement overrides an enterprise's negotiated zero-retention agreement for that traffic with no opt-out (source).

Stated safeguards: no training use and no non-safety use; all human access logged; deletion after 30 days "in almost all cases"; human review only via a controlled access path after an automated flag. Stated purpose: mitigating jailbreaks and reducing false positives in the safeguard layer — the same layer this page documents above, whose biology fallbacks were cut ~85% on 2026-08-07.

This belongs on the model page and not only on the policy page: it is a property of using this model, it is the reason a regulated buyer may be unable to deploy it, and it is not visible in the Spec table's Availability row. See Safety Monitoring and Data Retention for the contrast with OpenAI's opposite position, announced 2026-08-19.

Reported to be changing, 2026-08-20 — but not for this model, as far as anything read says. Bloomberg, citing an unnamed source, reports Anthropic will keep the 30-day requirement while letting enterprise customers hold the data on their own cloud infrastructure, rolling out later this year (source). Nothing read states whether Fable 5's total lack of ZDR support is affected — that is a separate provision from where the retained data sits, and the reporting addresses only the second. Recorded here because a regulated buyer's blocker on this model is the first provision, not the second, and it is unchanged by anything published so far.

Cyber Jailbreak Severity (CJS) Framework

Anthropic proposed the CJS framework — co-developed with Amazon, Microsoft, Google, and Glasswing partners — as an industry standard for rating AI jailbreak risk. Four axes: capability gain, breadth, ease of weaponization, discoverability. Five severity bands: CJS-0 (Informational) through CJS-4 (Critical). Also launched a HackerOne bug bounty for Fable 5 cyber jailbreaks. → AI-Enabled Cyberattacks (source) (Anthropic)

Compared To

Model	SWE-bench Pro	Tier	Public?
Claude Fable 5	80.3%	Mythos-class (gated)	✅ Yes
Claude Mythos Preview	77.8%	Mythos Preview	🔒 Glasswing only
Claude Opus 4.8	69.2%	Opus	✅ Yes
GPT-5.5	58.6%	—	✅ Yes
Gemini 3.1 Pro	54.2%	—	✅ Yes

Related

• Claude Mythos Preview — the withheld predecessor; Mythos 5 is the Glasswing-restricted successor sharing Fable 5's weights
• Claude Opus 4.8 — previous public frontier (2026-05-28, SWE-bench Pro 69.2%)
• Anthropic
• AI-Enabled Cyberattacks — Glasswing context for Mythos 5

Conflicting Reports

• Anthropic has published two figures for Fable 5's SWE-bench Pro, and this is not resolved. The Fable 5 launch announcement of 2026-06-09 gives 80.3% (source); the Opus 5 launch announcement of 2026-07-24 gives Fable 5 80.0% in its comparison table (source). Both are Anthropic's own announcements, so the schema's "official beats third-party" rule does not choose between them, and nothing in either says the benchmark was re-run. Each page therefore prints the figure from the announcement it is about, and both carry this note. Do not silently reconcile them — a later announcement restating an earlier number is not the same as a correction, and no source calls it one.

Sources

• Anthropic — Claude Fable 5 and Mythos 5 (source)
• VentureBeat
• CNBC
• The Decoder
• GitHub Copilot Changelog
• AWS Blog
• Anthropic — Models overview — spec figures verified 2026-07-27
• Anthropic — Improving Fable 5's Biology Safeguards (2026-08-07) → snapshot