# Claude Fable 5 Is Back: The Suspension, the Resolution & What Changes Now | Alher Tech > Claude Fable 5 is back: the US lifted the export controls on June 30, 2026 and Anthropic restored global access on July 1. The 19-day suspension explained, where the model is available now, the July 7 deadline, and how to build AI that survives losing a vendor. - Canonical page: https://alhertech.com/en/blog/claude-fable-5-suspended/ - Site: Alher Tech (custom software, AI agents and SEO engineering, https://alhertech.com/) - Contact: https://alhertech.com/en/contact/ --- Update (July 3, 2026): Claude Fable 5 is back. The US Commerce Department lifted the export controls on June 30, and Anthropic restored global access on July 1 across the Claude API, claude.ai, Claude Code and Claude Cowork, with AWS Bedrock, Google Cloud and Microsoft Foundry re-enabled the same day. See the new section below for the conditions that remain, and note the included-usage window on paid plans ends July 7. On June 9, 2026, Anthropic launched Claude Fable 5, the first publicly available model in its new Mythos class and, in the company's own framing, the most capable model it had ever released to the general public. Seventy-two hours later, it was gone. On the evening of Friday, June 12, a United States government export-control directive forced Anthropic to disable both Fable 5 and its sibling Claude Mythos 5: not in one country, but for every user on the planet. This was not a bug, an outage, or a billing dispute. It was a national-security order from the US Department of Commerce, delivered to CEO Dario Amodei at 5:21 PM ET, that Anthropic publicly disagreed with, and ultimately got reversed. On June 30, 2026, the Commerce Department lifted the controls, and on July 1 Fable 5 returned to users worldwide, 19 days after it went dark. Below is the most complete, sourced account we could assemble: what Fable 5 actually is, the exact sequence that led to the shutdown, what the 'jailbreak' at the center of it really was (it is more ironic than you would expect), how the model came back and under what conditions, and the architecture lesson every team shipping AI should take from it. ## The Timeline: Launch, Global Shutdown, and the Road Back The events unfolded in days, not weeks, and the resolution took less than three. Here is the full sequence pieced together from Anthropic's own statements and reporting by Fortune, TechCrunch, Axios, Politico, Semafor, CNBC and Forbes: - Apr 2026: Anthropic previews the Mythos class to a limited set of partners through Project Glasswing, but declines a public release, warning the class is unusually powerful. - Jun 9, 2026: Public launch of Claude Fable 5 and Mythos 5 across the Claude API, AWS Bedrock, Google Cloud, Microsoft Foundry and more. Fable is offered free through June 22 on Pro, Max, Team and Enterprise plans. - Thu, Jun 11: Amazon CEO Andy Jassy raises concerns with senior administration officials. Amazon researchers had used a chain of prompts to get the model to return restricted cyberattack information. - Fri, Jun 12 (day): The administration asks Anthropic to fix the issue or withdraw the model. Per White House AI adviser David Sacks, Amodei refuses. Treasury Secretary Scott Bessent reportedly tells Amodei 'he was making a bad decision.' Anthropic says it was given 90 minutes. - Fri, Jun 12, 5:21 PM ET: The Commerce Department sends Amodei an export-control directive naming Claude Fable 5 and Claude Mythos 5 specifically, citing national-security authorities. - Fri, Jun 12 (evening): Anthropic disables both models globally for all users. Claude Opus 4.8 and every other model stay fully online. - Fri, Jun 26: The US government approves restoring Claude Mythos 5 for a set of US organizations in Project Glasswing: the first partial reversal. - Tue, Jun 30: The Commerce Department lifts the export controls on both models. Anthropic announces the redeployment the same day. - Wed, Jul 1: Claude Fable 5 returns globally (Claude API, claude.ai, Claude Code, Claude Cowork) and is re-enabled on AWS Bedrock, Google Cloud and Microsoft Foundry. The directive barred access by any foreign national, whether inside or outside the United States, and including Anthropic's own foreign-born employees. Because the company could not filter users by nationality in real time, it chose to switch both models off for everyone rather than risk non-compliance. ## The Return: How Fable 5 Came Back Online On June 30, 2026, 18 days after the directive, the Commerce Department withdrew the export controls on Fable 5 and Mythos 5, removing the license requirement for both models. Anthropic announced the reversal in a post titled 'Redeploying Claude Fable 5' and restored access the next day, July 1: globally and all at once, not as a phased rollout. Fable 5 is live again on the Claude API, claude.ai (Pro, Max, Team and Enterprise plans), Claude Code and Claude Cowork, and back on AWS Bedrock, Google Cloud and Microsoft Foundry. Claude Mythos 5 followed a narrower path: the government approved restoring it on June 26, but only for a set of US organizations in Project Glasswing. It remains invitation-only, with no self-serve sign-up, and access for international Glasswing partners is still being coordinated with the government. - Included usage ends July 7: On Pro, Max, Team and select Enterprise plans, Fable 5 is included for up to 50% of weekly usage limits only through July 7, 2026. After that it moves to usage credits. This window replaces the original launch promo (which ran to June 22), and Anthropic has not said whether it will extend it. Note that Fable 5 burns through weekly limits faster than Opus 4.8. - Same price, unchanged specs: $10 per million input tokens and $50 per million output tokens, 1M-token context window by default, up to 128K output tokens per request. US-only inference is available as an option at 1.1x pricing. - The safeguards stay on: Blocking classifiers for cybersecurity and biology remain active, with refusal rates materially higher than on previous Claude models. On consumer surfaces many of those queries route automatically to Claude Opus 4.8; on the API a blocked request returns HTTP 200 with stop_reason 'refusal', and the Opus 4.8 fallback is something you configure yourself (server-side fallbacks beta or the SDK middleware). Prompt-stage refusals are not billed. - 30-day retention still applies: Both models remain 'Covered Models' with mandatory 30-day data retention and no zero-retention option. On AWS Bedrock this requires an explicit opt-in to provider_data_share via the Data Retention API: there is no console UI for it at launch. As part of the resolution, Anthropic committed to proactive risk detection, collaboration with the government on security protocols and standards, and prompt reporting of malicious activity. It is also proposing an industry-wide framework for scoring jailbreak severity together with Amazon, Microsoft, Google and other Project Glasswing partners. ## What the 'Jailbreak' Actually Was, and Why It's Ironic The trigger was a reported jailbreak of Fable 5's safeguards. But the details, in Anthropic's own words, are surprising. The company described it as 'a narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws.' That is the irony at the heart of this story. The Mythos class was built to be the strongest cybersecurity model in the world: its sibling, Mythos 5, is deployed to roughly 150 organizations in 15+ countries through Project Glasswing specifically to find and patch vulnerabilities in critical software. The very capability that makes it a world-class defender (reasoning through a codebase to surface exploitable flaws) is dual-use. Pointed the other way, it is exactly what a state-level attacker would want. Sacks said a 'highly credible, trusted partner of both Anthropic and the government' had identified the jailbreak in Fable 5's guardrails, and that after Amodei declined to fix or pull the model, the administration issued the export control 'reluctantly.' Semafor reported the government suspected a China-linked group may have already used the jailbreak Amazon discovered, though the evidence was described as unclear, and Anthropic disputes that the White House ever raised Chinese access in their conversations. Anthropic's broader position is that no model is perfectly jailbreak-proof: it states that 'perfect jailbreak resistance is not currently possible for any model provider,' while maintaining that Fable's safeguards are 'substantially more effective than those of any previously deployed model.' ## What Claude Fable 5 Actually Is Fable 5 and Mythos 5 share the same underlying model and published specifications; Fable is the public-facing variant with additional safeguards. For context on what's back online: - 1M-token context, 128K output: A 1 million-token context window by default, with up to 128,000 output tokens per request, and the ability to stay on a single task across millions of tokens without losing the thread. - Premium pricing: $10 per million input tokens and $50 per million output tokens, roughly double Claude Opus 4.8. Prompt-cache reads were listed at $1 per million tokens. - Built-in high-risk refusals: In domains like cybersecurity, biology and chemistry, Fable 5 was designed to block and defer to Claude Opus 4.8 instead, an unusual safety architecture baked into the product. - Mandatory 30-day retention: Both models required 30-day traffic retention for every user (even those previously on zero-retention agreements) as 'Covered Models.' ## Anthropic-reported benchmarks On the numbers Anthropic and early third-party testers published, Fable 5 sits at the frontier: - SWE-bench Verified: 95%. Real-world software-engineering task resolution. - SWE-bench Pro: 80%. Harder, agentic real-world coding. - Tau2-Bench: 98.5%. Tool use and agentic reliability. - GPQA Diamond: 92.6%. Graduate-level expert reasoning (science). - Hex analytics eval: ~90%. First model to clear 90% on Hex's core analytics benchmark. ## Fable 5 Is Back: What To Check Before You Re-Enable It Since July 1, calls to claude-fable-5 work again: same model ID, same pricing, same specs. But 'back online' does not mean 'back to normal'. Before you route production traffic to it again, check these four things: - If you migrated away during the outage: You can repoint to claude-fable-5 whenever you like: the model returned globally on July 1 with no waitlist. If your abstraction layer made the June switch painless, switching back is a one-line change; that was the whole point. - Keep your fallback chain in place: The cybersecurity and biology classifiers are still active, with materially higher refusal rates than other Claude models. Even benign adjacent work can occasionally trip them. Configure the Opus 4.8 fallback on the API (it is opt-in, not automatic) rather than removing the resilience you just built. - Re-check your data posture: The mandatory 30-day retention policy still applies to both models, with no zero-retention option. On AWS Bedrock, you must explicitly opt in to provider_data_share via the Data Retention API before requests will work. - Mind the July 7 deadline: If you're on a Pro, Max, Team or select Enterprise plan, the included-usage window (up to 50% of weekly limits) closes July 7, 2026. After that Fable 5 runs on usage credits. ## The Real Lesson: Never Bet Your Product on a Single Model The Fable 5 suspension is a textbook example of a risk most teams underestimate: model availability is not guaranteed. A model you depend on can vanish overnight: through deprecation, a price change, a capacity limit, or, as here, a government order entirely outside your control. And this time, the warning was 90 minutes. Teams that wired Fable 5 in as a single hardcoded dependency had production break with no real notice. Teams that built vendor-agnostic architectures barely noticed: a fallback engaged and traffic kept flowing. That difference is an engineering decision you make before the incident, not after it. - Abstract the model behind an interface: Never call a vendor SDK directly from your business logic. A thin routing layer lets you swap models in one place instead of across your whole codebase. - Build resilient fallback chains: Primary model down? Fail over automatically to a secondary (another Anthropic model, another vendor, or a self-hosted LLM) with retries and circuit breakers. - Keep a self-hosted option in reserve: For workloads where privacy or guaranteed availability matters, an open model on your own infrastructure can't be pulled by anyone's export order. - Handle refusals as a first-class case: Fable already deferred high-risk prompts to another model. Your system should treat refusals and outages the same way: with a defined fallback path, not a 500 error. ## How Alher Tech Builds AI That Doesn't Break We don't ship demos: we put AI agents and automations into production, and we design them so that no single model is a single point of failure. The Fable 5 suspension is exactly the scenario our architectures are built to absorb. - Vendor-agnostic model routing: Every request is routed to the right model behind a single abstraction: a frontier model for the hardest reasoning, cheaper models for the long tail, self-hosted LLMs when privacy or availability is the priority. - Resilient fallback chains: Cloud frontier model as primary, alternative providers or self-hosted models as fallback, with retries, circuit breakers and real-time alerts so the system always responds, even when a model disappears. - Private, self-hosted deployments: When you can't afford to depend on an external API, we run open models on your own GPUs. No export order, deprecation or rate limit can take them down. - Evals and guardrails baked in: We benchmark models head-to-head on your real tasks before committing, and wrap agents in sandboxes, allowlists and spend caps so they're safe to run in production. If the Fable 5 episode made you realize how exposed your AI stack is to a single vendor, that's exactly the problem we solve. Let's design an architecture that keeps working no matter which model is online tomorrow. ## Conclusion Claude Fable 5 was, by Anthropic's own account, one of the safest and most capable models ever deployed, and it was still pulled from the market three days after launch by a government order nobody could predict, over a 'jailbreak' that amounted to asking it to fix software flaws. Nineteen days later it was back, just as suddenly. That tension between capability and control is the whole story. The happy ending doesn't change the lesson. The teams hurt by this weren't the ones using Fable 5: they were the ones who could only use Fable 5. And the teams re-enabling it this week without friction are the ones whose architecture made the outage a routing change. In frontier AI, the durable advantage isn't picking the best model: it's building systems that survive losing it. If you're integrating frontier models into a product, an agent, or an automation, we can help you design the routing, fallbacks and self-hosted options that turn a vendor outage into a non-event. ## Related articles - Claude Opus 4.8: What's New, Benchmarks, Pricing & How to Use It - AI Agents Development: Production-Grade AI for Your Business - Software Development Life Cycle (SDLC) with AI: Complete Guide 2026 ## Frequently asked questions ### Is Claude Fable 5 available right now? Yes. The US Commerce Department lifted the export controls on June 30, 2026, and Anthropic restored global access on July 1: on the Claude API, claude.ai (Pro, Max, Team and Enterprise), Claude Code, Claude Cowork, AWS Bedrock, Google Cloud and Microsoft Foundry. ### Why was Claude Fable 5 suspended? The US Department of Commerce issued an export-control directive on June 12, 2026, citing national-security authorities, after a reported jailbreak of Fable 5's safeguards was flagged to the White House. Anthropic disabled the models globally because it could not restrict access by nationality in real time. ### How long was Claude Fable 5 offline? 19 days: from the evening of June 12 to July 1, 2026. The Commerce Department lifted the controls on June 30, and Anthropic redeployed the model globally, all at once, the next day. ### Until when is Fable 5 included in paid plans? Through July 7, 2026. On Pro, Max, Team and select Enterprise plans, Fable 5 is included for up to 50% of weekly usage limits until that date; afterwards it moves to usage credits at $10/$50 per million input/output tokens. ### What's the difference between Fable 5 and Mythos 5? They share the same underlying model. Fable 5 is the public-facing version with additional safeguards and is generally available again. Mythos 5 was restored earlier (June 26) but only for approved US organizations in Project Glasswing. It remains invitation-only, with international access still being coordinated with the government. ### Do the old conditions still apply? Yes. Both models keep the mandatory 30-day data-retention policy (no zero-retention option; on Bedrock it requires an explicit provider_data_share opt-in), and the cybersecurity and biology classifiers remain active with materially higher refusal rates than other Claude models. Pricing is unchanged at $10/$50 per million tokens.