Story perspectives

Claude Fable 5 Returns, Scores Plummet on BridgeBench

7/4/2026

Summary

Claude Fable5 Regression

Claude Fable 5 returned online with reduced performance compared with the July 1 release.
BridgeMind reran it on BridgeBench, recording score drops of debugging 86.2->25.9, refactoring 73.6->38.4, and hallucination-resistance 75.9->61.7.
The X post explained that debugging used a TypeScript repair benchmark, refactoring measured behavior-preservation and structural compliance, and hallucination tested false-claim detection across 30 tasks, six clusters, and 175 questions.

Summary

Anthropic Blocks Access

An export-control order, prompted by an Amazon report of a safeguard bypass, forced Anthropic to block foreign nationals from Claude Fable 5.
Because Anthropic cannot verify nationality in real time, it disabled both Claude models.
Pro, Max, Team and Enterprise customers receive Fable 5 at 50 % of limits until July 7, after which access shifts to usage credits.
Fable 5 will resume on July 1 across Claude Platform, Claude.ai, Claude Code and Claude Cowork.

Summary

Anthropic Blocks Exploit

Anthropic introduced an upgraded safety classifier for Claude Fable 5 that blocks the exploit technique in more than 99% of attempts.
The classifier was trained specifically to target the behavior that raised export-control concerns.
When the classifier rejects a request, the system notifies the user and forwards it to Claude Opus 4.8.
The fallback to Claude Opus 4.8 maintains service continuity while preserving the safety barrier without extra user action.

History

Historical Context

In 1999 the U.S. Department of Commerce amended the Export Administration Regulations to permit 56-bit encryption and 1024-bit RSA without licenses, a shift from earlier strict controls. That precedent guided Anthropic’s July 2024 export-control order, which bars foreign nationals from accessing the Claude Fable 5 model and caps usage at half capacity. The episode shows how decades-old export rules still shape AI deployment.

1 of 4

Claude Fable5 Regression

Claude Fable 5 returned online with reduced performance compared with the July 1 release.
BridgeMind reran it on BridgeBench, recording score drops of debugging 86.2->25.9, refactoring 73.6->38.4, and hallucination-resistance 75.9->61.7.
The X post explained that debugging used a TypeScript repair benchmark, refactoring measured behavior-preservation and structural compliance, and hallucination tested false-claim detection across 30 tasks, six clusters, and 175 questions.

1 / 4