Summary
- BridgeBench's assessment of Claude Fable 5's performance plummeted from 86.2 to 25.9 following its July 1 return, attributed to the safety classifier rerouting most tasks to Opus 4.8 rather than a decline in the model's capabilities.
- Arena.AI conducted a series of blind human-preference surveys and reported that Fable 5's performance remained consistent with its June version, with improvements noted in specific areas such as document and expert text.
- Anthropic acknowledged that its new classifiers may generate false positives during routine coding and debugging tasks, promising future refinements but without a specific timeline for updates.
When Claude Fable 5 resumed operations on July 1, feedback on social media was overwhelmingly negative, labeling it as broken, nerfed, lobotomized, and underperforming, suggesting it was not the same model anymore.
I've been using Fable 5 all day, continuing my work with Opus.
The reports are accurate.
It feels completely nerfed.
Politics has once again stifled technological progress for civilians. https://t.co/Ed3jrqOxbK
— BharadwajC (@bwjbuild) July 2, 2026
User backlash was significant. Simultaneously, two evaluation platforms—BridgeBench AI and Arena AI—released contrasting findings. One indicated a drastic decline in output quality, while the other noted minimal changes that might not be easily noticeable.
Both evaluations have valid points.
In essence, the model’s intelligence hasn't diminished; rather, the gatekeeper mechanism has become more stringent. This distinction is crucial depending on the intended use of Fable.
BridgeBench's Assessment
BridgeMind, an AI evaluation service, conducted a full coding suite test on the July 1 version of Fable 5 immediately after its re-launch.
BridgeBench evaluates real-world coding tasks across various categories such as debugging, refactoring, and hallucination resistance, scoring from 0 to 100 based on task completion. The results were alarming: Debugging scores dropped from 86.2 to 25.9, Refactoring from 73.6 to 38.4, and Hallucination resistance from 75.9 to 61.7.
FABLE 5 RETURNED NERFED.
We re-evaluated the July 1st Claude Fable 5 on BridgeBench.
The findings are harsh:
Debugging: 86.2 → 25.9
Refactoring: 73.6 → 38.4
Hallucination: 75.9 → 61.7The new guardrails are excessively restrictive across many tasks, reverting to Opus… pic.twitter.com/tcUDDXpZMF
— BridgeMind (@bridgemindai) July 2, 2026
This classifier was implemented as a condition for Fable's return, designed to prevent the jailbreak technique reported by Amazon, which allowed Fable 5 to pinpoint and showcase software vulnerabilities. While effective, it also misidentifies many tasks it shouldn't, categorizing debugging TypeScript as "security work" and triggering frequent rerouting.
Arena.AI's Findings
Arena.AI, a benchmarking platform for LLMs, approached the assessment differently. It gathers thousands of blind human-preference votes across various categories—including text, vision, document, code, and agent—and utilizes Elo scoring, a chess-based rating system that adjusts for statistical variability across head-to-head comparisons. This method reflects perceived quality rather than infrastructure routing.
The community has been curious about how Claude Fable 5 performs pre- and post-deployment.
We gathered thousands of votes on the new endpoint across various Arenas—Text, Vision, Document, Code, and Agent—and here’s an initial score overview.
So far, scores appear largely… https://t.co/FKDaPpz10e pic.twitter.com/1nJDHqnlIj
— Arena.ai (@arena) July 2, 2026
The comparison indicated that Fable 5 mostly maintained its performance. Frontend coding saw a slight drop from 1650 to 1623 Elo—a change noted by Arena to be within the confidence interval as more data accumulates. Document performance improved by 34 points, expert text by 25, and creative writing increased slightly by 9. The areas that saw declines were coding at -18 and hard prompts at -3—tasks most likely intercepted by the classifier before Fable could respond.
In summary, when Fable 5 is able to engage with the task directly, it performs consistently. The dissatisfaction expressed on X stems not from a decline in model capability but from the frequent rerouting of tasks away from Fable.
Who is Impacted and Who is Not
General users engaged in creative writing, document analysis, research, and expert-level text inquiries may notice minimal to no changes. These categories show stable or improved performance according to Arena.AI. Any improvements might be too subtle to detect, especially in subjective tasks like creative writing where results are difficult to quantify.
Essentially, writers, researchers, and analysts will receive the expected performance from Fable 5. However, developers face a different scenario.
Individuals working in areas related to security—such as memory management coding or tasks involving terms like "vulnerability," "exploit," "hook," or even "fix"—will frequently encounter rerouting.
The disparity between BridgeBench's significant drop and Arena's consistent scores can be attributed to the nature of the tasks. BridgeBench's suite consists of prompts that typically activate the new classifier, while Arena's human voters pose a broader range of queries, many of which do not appear as exploit code to the safety layer.
Anthropic has indicated that the classifiers will evolve, admitting they currently have an overly broad scope. The initial restriction was implemented after Amazon researchers discovered a method for Fable to identify and illustrate software vulnerabilities, which the U.S. government classified as a national security issue. The solution involved making the classifier sufficiently cautious to capture that and related tasks, with plans to fine-tune later.
However, no specific date has been provided for this adjustment.