Claude Fable 5 returned online July 1, and two AI benchmarking platforms published conflicting performance assessments the same day. BridgeBench reported Claude Fable 5's debugging score collapsed from 86.2 to 25.9 after reinstatement, while Arena.AI found performance largely unchanged through thousands of blind human-preference votes. The divergence stems from Anthropic's new safety classifier routing most coding and debugging tasks to Claude Opus 4.8 rather than allowing Fable 5 to handle them directly. Anthropic has acknowledged the classifier produces false positives on routine coding tasks. The company deployed the conservative classifier as a condition of Fable 5's reinstatement following a security vulnerability demonstration reported by Amazon researchers.
BridgeBench Records Severe Score Drops Across Coding Categories
BridgeMind re-ran its full coding suite against the July 1 version of Fable 5 the day it came back. BridgeBench tests real-world coding tasks across categories including debugging, refactoring, and hallucination resistance, scored 0–100 on how well the model completes each category. Debugging fell from 86.2 to 25.9, Refactoring from 73.6 to 38.4, and Hallucination resistance from 75.9 to 61.7.
Of 12 TypeScript debugging tasks, only three actually reached Fable 5. The remaining nine were intercepted by Anthropic's new safety classifier and rerouted to Claude Opus 4.8. BridgeBench scores every fallback as zero, because the model that answered wasn't the one under evaluation. The classifier was trained to block the Amazon-reported jailbreak technique that got Fable 5 to identify and demonstrate software vulnerabilities. Debugging TypeScript looks enough like security work to the classifier that the fallback fires constantly.
Arena.AI Human Preference Votes Show Flat to Improved Performance
Arena.AI ran the same question through a different lens. The platform collects thousands of blind human-preference votes across multiple categories—text, vision, document, code, and agent—and ranks models using Elo scoring. When two models go head-to-head anonymously and humans pick a winner, the score reflects actual perceived quality, not infrastructure routing.
The before-and-after comparison showed Fable 5 largely holding its ground. Frontend code dropped from 1650 to 1623 Elo—a difference Arena noted is within the confidence interval as data keeps accumulating. Document performance improved by 34 points. Expert text went up 25. Creative writing edged up slightly by 9. The categories that declined—Coding at -18, hard prompts at -3—are precisely where the classifier is most likely to intercept the prompt before Fable can answer.
When Fable 5 actually handles the task, it still performs like Fable 5. General users doing creative writing, document analysis, research, and expert-level text queries will likely notice little to no difference. Those are the categories where Arena.AI shows flat or improved performance. Developers working in security-adjacent territory—coding memory management, anything touching words like vulnerability, exploit, hook, or fix—will hit the fallback regularly.
Anthropic Acknowledges False Positives on Routine Coding Work
Anthropic has said the classifiers will improve over time, acknowledging they currently cast too wide a net. The original ban came after Amazon researchers found a technique to get Fable to identify and demonstrate software vulnerabilities, and the U.S. government treated that as a national security threat. The fix was to make the classifier conservative enough to catch that and everything around it, then tune it down later. Anthropic has given no target date for when that will happen.
FAQ
What caused Claude Fable 5's debugging score to drop from 86.2 to 25.9 after July 1?
The drop resulted from Anthropic's new safety classifier routing nine of twelve debugging tasks to Claude Opus 4.8 instead of allowing Fable 5 to handle them. BridgeBench scores every fallback as zero because the evaluated model did not answer. The classifier was deployed to block the Amazon-reported jailbreak technique that got Fable 5 to demonstrate software vulnerabilities.
How did Arena.AI's human preference testing differ from BridgeBench's results?
Arena.AI collected thousands of blind human-preference votes across text, vision, document, code, and agent categories. The platform found Fable 5 performance mostly flat compared to the June version, with document performance improving by 34 points and expert text up 25 points. Frontend code dropped from 1650 to 1623 Elo, a difference Arena noted is within the confidence interval.
When will Anthropic refine the safety classifier to reduce false positives?
Anthropic has acknowledged the new classifiers produce false positives on routine coding and debugging tasks and stated the system will be refined over time. The company has given no timeline for when refinements will occur.
