After the United States Department of Commerce lifted export restrictions on two new Anthropic models, Claude Fable 5 will resume its global visit on Wednesday, covering Claude.ai, Claude Platform, Claude Code and Claude Work. Previously, this restriction had led to a global down-line model for companies starting on 12 June.
Mythos 5 is still not fully recovered
According to Anthropic, the more capable Claude Mythos 5 is still open only to approved partners and is not resuming access to all users.
The escape was based on an escape test report. Amazon researchers found that the Fable 5 security fence could be bypassed, that models would identify software gaps and, in individual cases, generate codes that could be used to exploit them.
Since the export restrictions then applied to foreign citizens on a global scale, and Anthony could not verify the nationality of the users in real time on a case-by-case basis, the company eventually chose to discontinue the two models for all users in order to avoid the risk of non-compliance.
The new security classifier intercepts over 99%.
In its note, Anthropic stated that the company had retrained a safety taxonomy device specifically to identify the method of escape for this disclosure. According to them, the new system can intercept more than 99 per cent of related attempts.
If the Fable 5 request is systematically marked, the Platform will automatically switch the request to Claude Opus 4.8 and send a reminder to the user. The AI Standards and Innovation Centre of the United States Department of Commerce, which reviewed both old and new protections, considered them “very strong”. However, Anthropic also admits that stricter filtering can lead to more misinformation.
Cloud deployment resumed later
In terms of access arrangements, Pro, Max, Team and some Enterprise users can use the Fable 5 service, which accounts for up to half of the weekly amount, by July 7. Thereafter, the format of the visit will be changed to a utilization factor.
Anthropic also states that availability on AWS, Google Cloud and Microsoft Foundation will be restored “as soon as possible”, which means that business clients relying on cloud-based hosting for deployment will recover later than Claude owns.
During the decommissioning of the two models, competitors took the opportunity to advance the relevant products. According to the report, OpenAI GPT-5.5-Cyber exceeded Mythos 5 83.8 per cent in the CyberGym benchmark test at University of California Berkeley by 85.6 per cent.
Anthropic proposes to establish a classification standard with a technology company
Anthropic also indicated that a classification framework for the severity of prison escape was being drafted with Amazon, Microsoft and Google to assess the dimensions of capacity enhancement, scope of application, weaponization difficulties and detectability.
At the same time, the company promised to provide the model and its security measures to the Government in the future for testing before the forward model was publicly released. Anthropic is also on line with the new HackerOne project to collect Fable 5 bypass cases submitted by researchers.