The news everyone was waiting for is here! Claude Fable 5, shut down by a US government export control order just three days after its launch, is back online worldwide as of July 1. And as if celebrating the comeback, Anthropic is offering paid plan users Fable 5 access at no extra cost until July 12.

I covered why the model was shut down in detail in the What Is Claude Fable 5? post. Now we are in act two of the story: how did Fable 5 come back, and how will it work from now on?

What Happened? A Quick Recap 📋

A quick refresher:

  • June 9: Fable 5 and Mythos 5 launched.
  • June 12: Citing a jailbreak method found by Amazon researchers (getting the model to identify software vulnerabilities and write exploit code), the US government applied export controls. Since nationality could not be verified in real time, Anthropic shut both models down globally.
  • June 26: Mythos 5 was restored for roughly 100 US organizations defending critical infrastructure.
  • June 30: Export controls were fully lifted with the approval of Commerce Secretary Howard Lutnick.
  • July 1: Fable 5 returned worldwide.

Anthropic’s testing also revealed an interesting point: the same jailbreak technique produced identical results on weaker models like Opus 4.8, GPT-5.5, and Kimi K2.7. In other words, the exploit did not expose any dangerous capability unique to Fable 5.

How Will Fable 5 Work From Now On? 🛡️

The real price of the comeback is a new safety layer. Anthropic developed a dedicated safety classifier targeting the exact bypass technique reported by Amazon. Here is how the system works:

  1. Over 99% blocking: The new classifier catches the reported jailbreak method in more than 99% of cases.
  2. Redirection to Opus 4.8: Instead of being silently refused, blocked requests are handed off to Claude Opus 4.8, and the user is notified about the redirection.
  3. Wide safety margin: Anthropic deliberately tuned the classifier with a large safety margin. The trade-off is that some legitimate requests get blocked by mistake.
Fable 5 False Positive Warning: Coding Requests May Be Affected
Due to the wide safety margin, legitimate coding requests that look like security analysis, debugging, or exploit work may occasionally be routed to Opus 4.8. If your request gets redirected, that is by design, not a bug. On the API side, the response returns as HTTP 200 with stop_reason: "refusal".

Beyond that, Anthropic expanded its “defense in depth” approach: refusal behavior at the training stage, retroactive analysis of usage patterns, and automated classifiers all operate as overlapping layers.

Was Fable 5 Nerfed? Why Are Users Complaining? 📉

Just days after the comeback, social media and developer forums filled up with the same question: “Is Fable 5 weaker than before?” The complaints are concrete: constant fallbacks to Opus 4.8 even on routine tasks, blocks on systems-level C/C++/Rust code, and in one user’s words, “it didn’t even let me search for dead code without switching to Opus.”

The numbers fed that perception too. Independent coding evaluation group BridgeBench reported striking drops after the redeployment:

  • Debugging: 86.2 → 25.9 (roughly a 70% drop)
  • Refactoring: 73.6 → 38.4

So did the model actually get weaker? The truth is far more interesting. A look at BridgeBench’s methodology reveals the real culprit: out of 12 TypeScript debugging tasks, only 3 actually reached Fable 5. The remaining 9 were rerouted to Opus 4.8 by the new safety classifier, and the benchmark scored those fallbacks as zero. In other words, the test measured the wrong model on most questions.

Arena.AI Blind Tests: Model Weights Unchanged
Data from Arena.AI, which runs thousands of blind human preference comparisons, paints a different picture: overall performance between versions is nearly flat. There are even gains: +34 Elo on document work, +25 Elo on expert text, +9 Elo on creative writing. The only decline is -18 Elo in coding, exactly where the classifier intervenes most aggressively. Conclusion: the model itself was not nerfed, the router is just paranoid.

The impact varies dramatically by use case:

  • Writers, researchers, analysts: The Fable 5 you expected is right there, even better in some areas.
  • Developers: Memory management, Win32 API work, and files containing words like “security”, “vulnerable”, “unsafe”, or “hook” regularly trip the classifier and get the task dropped to Opus 4.8.

Anthropic has not yet publicly responded to the false positive reports, only stating that the classifiers “will improve over time”. No target date. In short, it is worth checking which model actually answered for the price you paid: Claude explicitly notifies you whenever a redirection happens.

New Industry-Wide Steps 🤝

This incident did not end with just a model update, it kicked off lasting changes across the industry:

  • Jailbreak severity framework: Together with Amazon, Microsoft, and Google, Anthropic developed shared criteria for assessing jailbreak severity. There are four dimensions: capability gain, breadth, ease of weaponization, and discoverability.
  • HackerOne bounty program: A new bug bounty program was launched specifically for cybersecurity jailbreaks in Fable 5. Submissions are monitored 24/7, and fixes for critical cases ship as soon as they are confirmed.
  • Government collaboration: New commitments include pre-release access to national security-relevant models, independent evaluation, and rapid sharing of jailbreak information.

Access and Pricing: The July 12 Opportunity 💰

Fable 5 is now available worldwide through Claude.ai, Claude Platform, Claude Code, and Claude Cowork. Access via AWS, Google Cloud, and Microsoft Foundry is being restored quickly as well.

Claude Fable 5 Free Usage: Until July 12
Pro, Max, Team, and Enterprise subscribers can use Fable 5 for up to 50% of their weekly usage limits at no additional cost. The promotion was originally set to end on July 7 but has been extended until July 12, 2026. After that, usage continues via usage credits. If you want to try it, now is the time!

API pricing has not changed: $10 input / $50 output per million tokens, with the 90% input discount for prompt caching still in place. The situation is different for Mythos 5: the unrestricted version remains available only to roughly 100 US organizations defending critical infrastructure, with negotiations for broader access ongoing.

Conclusion: From the First Government Takedown to the First Comeback 🎯

The Fable 5 saga made AI history once already: a frontier model pulled from the market not by its maker but by a government order. Now we have witnessed a second first: the same model returning with hardened safety layers and new industry-wide standards. This cycle may become the template for how powerful model launches unfold from here on.

So, will you try Fable 5 during the included usage window until July 12? What do you think about the redirection system? Let’s meet in the comments! 👇🏻

AI-Generated Content Notice
This blog post was entirely generated by artificial intelligence. While AI helps create content, it may still contain errors or biases. Verify critical details before using them.