Anthropic's Flagship Model Draws User Criticism Over Tightened Safety Classifiers

Anthropic's Claude Fable 5 is facing mounting user criticism following its re-release on July 1, with developers reporting sharp declines in coding, debugging, and agentic task performance. The company attributes the change not to the model itself, but to stricter safety classifiers deployed after a brief government-mandated suspension.

Benchmark Scores Drop Sharply After Re-Release

Benchmark group BridgeMind re-ran its BridgeBench suite on the July 1 version of Fable 5 and recorded significant score drops across key categories. Debugging fell from 86.2 to 25.9, refactoring declined from 73.6 to 38.4, and hallucination handling slipped from 75.9 to 61.7.

The mechanics behind the numbers are important: only 3 of 12 debugging tasks completed without falling back to Claude Opus 4.8. Every fallback instance scored zero, meaning the collapse reflects blocked tasks rather than degraded reasoning. BridgeMind noted that when Fable 5 does complete a task, its output matches its June performance.

'The model did not get worse. It got caged,' BridgeMind stated in its report.

Timeline: Suspension, Export Controls, and Restored Access

Anthropic originally launched Fable 5 on June 9. Washington pulled it offline three days later. Export controls were lifted on June 30 — four days after Mythos 5 access was restored for approximately 100 US institutions. The re-release on July 1 came with usage restrictions:

• Fable 5 draws from only 50% of weekly usage caps through July 7
• After July 7, usage shifts to paid usage credits
• Blocked requests are automatically routed to Opus 4.8, with users receiving a notification

Anthropic Defends Its Safety Approach

In a June 30 statement, Anthropic acknowledged it deliberately widened its safety margin, meaning classifiers now block requests that are probably benign. The company conceded the filter flags more legitimate coding and debugging work than before.

An improved filter blocks a known bypass technique in over 99% of attempts, according to Amazon researchers. Anthropic also noted that its own internal tests showed Fable 5 posed no unique cybersecurity risk — rival models, including GPT-5.5 and Kimi K2.7, identified the same vulnerabilities. US Commerce Department researchers tested both safeguard versions and judged them extraordinarily strong.

Broader Industry Implications

The episode carries consequences beyond a single product cycle. The suspension prompted Europe to actively court Anthropic for partnerships, while Chinese AI models continue to gain ground on US frontier labs.

Anthropic is currently drafting a jailbreak severity framework in collaboration with Amazon, Microsoft, and Google. Whether the classifiers can shed false positives quickly is expected to determine whether power users remain on the platform or migrate to competitors.