Claude Fable 5 is back online, but the version developers are testing may not behave like the model that made Anthropic’s latest release a flashpoint for Washington.
Community-led AI platform BridgeMind said a rerun of the July 1 version of Claude Fable 5 on BridgeBench showed sharp declines across several coding-related measures, with debugging falling from 86.2 to 25.9, refactoring from 73.6 to 38.4, and hallucination resistance from 75.9 to 61.7.
The account shared on X a BridgeBench debugging as a TypeScript repair benchmark, refactoring as a behavior-preservation and structural-compliance test, and hallucination as a code-analysis false-claim test across 30 tasks, six clusters, and 175 questions.
Anthropic said access to Fable 5 and Mythos 5 was restored after the Commerce Department lifted export controls that had applied to the models. The company said the order required it to restrict access by foreign nationals, including foreign nationals inside the US, and that it disabled both models for all users because it had no reliable way to verify nationality in real time.
Sponsored · Selkirk Copper Mines
The company said Fable 5 would return July 1 on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. It also said Pro, Max, Team, and select Enterprise users would get Fable 5 within up to 50% of weekly usage limits through July 7, after which access would shift to usage credits.
The operational tension is buried in Anthropic’s own explanation. The company said the export-control order followed an Amazon researcher report involving a method that bypassed Fable 5 safeguards and led the model to identify software vulnerabilities. In one case, Anthropic said the model produced code showing how a vulnerability could be exploited.
Anthropic’s fix was not simply to bring the model back. It trained an improved safety classifier aimed at the reported behavior. When Fable 5 blocks a request, Anthropic said the user is notified and the request is instead sent to Claude Opus 4.8. The company said the new classifier blocks the specific reported technique in more than 99% of cases.
That is the same mechanism BridgeMind is now criticizing. It said the “new guardrails are kicking in on way too many tasks and falling back to Opus 4.8,” adding that when tasks can be completed, Fable 5 appears to function as it did on June 12.
“The problem is that the guardrails are too strict,” it said.
Anthropic had already warned this could happen. In its redeployment note, the company said the classifier update “comes at the cost of flagging benign requests more often during routine coding and debugging tasks,” while saying it would refine the system to better separate misuse from legitimate work.
The dispute began as a national-security intervention, not a product-quality fight. Reuters reported that the Commerce Department lifted export restrictions on Fable 5 and Mythos 5 less than three weeks after requiring Anthropic to suspend access over security risks. It also reported that Anthropic is working with Amazon, Microsoft, Google parent Alphabet, and other Glasswing partners on common standards for assessing and fixing AI jailbreaks.
Commerce’s reversal was not unconditional in spirit. Commerce Secretary Howard Lutnick’s letter said the department reserved the right to reevaluate its decision and reimpose licensing requirements if circumstances change or Anthropic fails to meet its commitments.
WIRED reported that the Trump administration lifted the controls after Anthropic agreed to extend a guardrail that blocks attempts to access restricted capabilities and routes those queries to Opus 4.8. Commerce’s Center for AI Standards and Innovation cleared the model after deciding its safeguards were robust enough for now.
That makes BridgeMind’s finding commercially awkward. The very mechanism that helped reopen Fable 5 may be the mechanism reducing its usefulness in real coding workflows. Safety compliance solved the access crisis, but it may have created a product-consistency crisis.
The broader policy issue is also unresolved. CSIS noted that the June 12 restriction required a Bureau of Industry and Security license for foreign persons to access the models, including foreign nationals inside or outside the US. That framework, built around export controls and deemed exports, becomes harder to apply cleanly when the product is a cloud-hosted model accessed by global users rather than a physical shipment.
Anthropic has argued that Fable 5 and Mythos 5 share the same underlying model, with Fable 5 released under stronger safeguards for general use and Mythos 5 limited to trusted defensive-cybersecurity partners. It also said Fable 5 does not provide unique offensive cyber capabilities because of those safeguards.
The market question now is narrower and harsher: can Anthropic tune the guardrails without reopening the political fight that got Fable 5 restricted in the first place?
