Fable 5 returns, but has it been lobotomized?

Disclaimer - This article was partially edited by Fable 5

What Happened

Rewind exactly a month back to June 9, 2026. Anthropic released Claude Fable 5 - its most capable public model ever, a Mythos-class system priced at roughly double Opus rates ($10/M input, $50/M output tokens). Early reviews praised its state-of-the-art performance across software engineering (e.g., leading SWE-Bench Pro), knowledge work, vision, and agentic tasks. Then, on June 12, it vanished. Amazon researchers reportedly found a way to bypass Fable 5's safeguards, prompting it to identify software vulnerabilities and, in at least one case, demonstrate exploit code. They escalated the finding to the U.S. government rather than Anthropic directly. The Commerce Department issued an export-control directive barring foreign nationals (including Anthropic's own non-U.S. staff) from accessing Fable 5 and the more powerful Mythos 5. With no practical way to enforce nationality checks in real time, Anthropic shut both models down globally.

On July 1, Fable 5 returned with the same underlying weights. The reinstatement condition: a new, aggressive safety classifier developed in coordination with the government. It now screens every request and all content the model reads (files, search results, connectors) across key zones - offensive cyber operations, most biology/chemistry work, distillation attacks, and certain frontier LLM development topics. Flagged items reroute to Claude Opus 4.8. Anthropic warned upfront that routine coding and debugging would sometimes trigger this fallback in the short term.

What others are saying

The immediate reaction on X was harsh: "nerfed," "broken," "lobotomized." The loudest data point came from BridgeMind (@bridgemindai), whose post showed BridgeBench scores cratering—debugging 86.2 → 25.9, refactoring 73.6 → 38.4.

Important context on the methodology: Of 12 TypeScript debugging tasks, only 3 actually reached Fable 5. The rest were intercepted by the classifier, routed to Opus 4.8, and scored as zeros (since the evaluator was testing Fable specifically). Debugging broken code often resembles vulnerability analysis, reliably tripping the wire. BridgeMind has a track record here; it previously called Opus 4.6 "nerfed" in a sampling artifact that Community Notes later highlighted.

But another post which followed said it might be the most impressive model they ahve ever seen.

Developer Morgan Linton (@morganlinton) provided a counter-narrative with practical benchmarks: Fable 5 on Low effort matched or beat GLM 5.2 High and Sonnet 5 High across 10 real post-cutoff engineering tasks (all ~80% success). Costs: $2.27 for Fable vs. $8.41 GLM and $15.81 Sonnet. He ran Fable non-stop on the $100/month Max plan without limits or guardrail hits for 95% of his Low/Medium coding work, praising its planning capabilities.

My own Fable 5 experience

Over the past few days I put Fable 5 to work across a pretty ordinary consulting workload: optimizing my own site for LLM/AI-engine discovery (AEO), syncing LinkedIn newsletter editions onto that site, sorting 200 Salesforce prospects to flag which leads were actually worth chasing, standing up a staging clone of my AWS production ECS/EC2 environment, and drafting a product roadmap for a security scanner I'm building. Most of that sailed through untouched - the LinkedIn sync, the lead scoring, the AWS staging deploy all ran on Fable start to finish. The one place I got caught was the scanner roadmap. The moment the conversation turned to what the product actually scans for, Fable 5 quietly dropped me to Opus 4.8 mid-task. That tracks with what others are reporting: the classifier isn't weighing intent, it's pattern-matching vocabulary, and "security scanner" sits close enough to the restricted zone to get flagged even for pure product-planning work with no exploit code anywhere near it.

Overall, Fable 5 was rock solid - especially on the website redesign and the AWS EC2 and ECS deployments.

🫤 Dileep's Skeptical Takeaway(s):

1. Fable 5 seems weirdly more assertive (human-like) - I experienced what many others are saying. The model just seems more confident and seamless. Let it use its own judgment as much as possible.
2. Effort level is the untold story. Many complaints about cost/limits stem from defaulting to High effort on tasks solvable at Low/Medium. Fable's efficiency shines when used thoughtfully.
3. It is not lobotomized, but Fable 5 basically does not engage and downgrades to Opus 4.8 if it is unsure. Builders working on anything security-adjacent or Biology adjacent —scanners, defensive tooling, even a product roadmap that mentions vulnerabilities—get caught by the same broad classifier as offensive researchers. Attackers can switch providers.
4. Fable 5 is going away today, so use it while you can. Free access phases out soon, shifting to full-rate usage. Winners are learning prompt scoping, monitoring fallbacks, using appropriate effort levels, and measuring cost-per-task - not raw tokens. Also, use it to write Skills and guidelines for Opus while it is still available.
5. You don't need Fable 5 for everything - It eats up a lot more tokens so Fable 5 should be used for the most complex of tasks. I did not notice much difference in creating solution briefs and marketing collateral.

Bottom line: Fable 5 remains one of the most powerful tools available when you work with (not against) its new realities.