Following the U.S. government’s decision to lift export controls on Anthropic’s Fable 5 model, the artificial intelligence company announced its impending return on Wednesday, July 1. This reinstatement comes after a period of restriction imposed by the Commerce Department due to concerns over a "jailbreak" or, as Anthropic prefers to term it, a bypass, that allowed the model to identify and demonstrate software vulnerabilities. Anthropic has now provided a detailed roadmap for Fable 5’s reintroduction, outlining its phased rollout, updated accessibility, and a clearer explanation of the incident that led to its temporary withdrawal.
The re-deployment strategy for Fable 5 signifies a significant step in Anthropic’s commitment to balancing cutting-edge AI capabilities with robust safety protocols. The company’s proactive engagement with government agencies and industry partners underscores a growing trend in the AI sector towards greater transparency and accountability in the development and deployment of powerful AI models.
Phased Rollout and Subscription Adjustments
Starting July 1, Fable 5 will be accessible across Anthropic’s Claude Platform, Claude.ai, Claude Code, and Claude Cowork. This availability will extend to users globally who subscribe to the Pro, Max, Team, and Enterprise plans. However, the initial access period will be subject to certain limitations. For subscription users, Fable 5 will be available for up to 50% of their weekly usage limits until July 7. Following this initial grace period, access to Fable 5 will transition to usage credits, priced in line with Anthropic’s API plans.
For standard Enterprise users, Fable 5 will not be included in their regular subscription allowance from the outset. Access for this user segment will be billed through usage credits immediately upon reintroduction. A brief grace period will be offered for premium Enterprise seats, granting them access through their subscription plan until July 7, after which they too will be subject to usage-based billing.
This tiered access strategy reflects Anthropic’s approach to managing the demand and potential impact of Fable 5 as it returns to the market. The initial limited access aims to allow for monitoring and further refinement of safety mechanisms before broader, unrestricted availability.
Originally, Fable 5 was slated for free access from June 9 to June 22, a period that was cut short by the imposition of export controls. Developers who previously accessed Fable 5 through major cloud platforms such as AWS, Google Cloud, and Microsoft Foundry can anticipate its re-enablement on those platforms in the near future.
Prior to its withdrawal, Anthropic’s pricing for Fable 5 was set at $10 per million input tokens and $50 per million output tokens. There is no indication that these pricing structures will undergo any changes with the model’s return.
Unpacking the "Jailbreak" Incident
Anthropic has shed more light on the circumstances surrounding the export controls. The company confirmed that researchers from Amazon discovered a method to bypass Fable 5’s safety protocols. This bypass involved prompting the model in a way that led it to identify specific software vulnerabilities. In at least one documented instance, the model was reportedly prompted to demonstrate how a discovered vulnerability could be exploited.
Crucially, Anthropic has emphasized that this particular technique did not reveal any unique or advanced "Mythos-level" cyber capabilities. Instead, the exploit was described as operating just below the threshold that would typically trigger Fable 5’s built-in safeguards.
In a significant finding, Anthropic’s investigation, conducted in collaboration with government agencies and partners including Amazon, revealed that other advanced AI models were also capable of identifying the same vulnerabilities. Models such as Claude Opus 4.8, GPT-5.5, and Kimi K2.7 were found to possess similar discovery capabilities. Furthermore, even more basic models, including Anthropic’s own Claude Haiku 4.5, were able to discern how to exploit the identified vulnerability. This suggests that the issue was not an isolated flaw in Fable 5 but rather a characteristic of advanced AI models’ ability to analyze complex code and identify potential weaknesses.
Enhanced Safety Classifiers and Broader Implications
To address the vulnerabilities and prevent future occurrences, Anthropic has developed and implemented an improved safety classifier. This system is designed to detect and block requests that could lead to the generation of harmful outputs. According to Anthropic, the newly implemented classifier effectively blocks the specific technique detailed in the Amazon report in over 99% of cases.
This enhancement involves a recalibration of the model’s safety thresholds. Anthropic has "deliberately set the safety classifiers to trigger on a set of requests that we know are likely benign" to establish a wider safety margin. This approach means that Fable 5 will now err on the side of caution, potentially blocking a greater number of requests that are not intended for malicious purposes.
This recalibration comes with an anticipated trade-off: an increase in the flagging of benign requests, particularly during routine coding and debugging tasks. Anthropic acknowledges this potential side effect, stating, "The new classifier also comes at the cost of flagging benign requests more often during routine coding and debugging tasks." The company has pledged to continue refining these safeguards to better differentiate between genuine misuse and legitimate user activity, aiming to minimize false positives.
The decision to focus these enhanced safety measures solely on Fable 5, despite other models exhibiting similar behaviors, appears to be a joint decision with the U.S. government. This targeted approach suggests a specific concern related to Fable 5’s capabilities or its perceived risk profile at the time of the incident.
For problematic requests that still manage to bypass the enhanced guardrails, Fable 5 will continue to route them to Opus 4.8. This is notable, as Anthropic has indicated that Opus 4.8 was also capable of replicating Fable 5’s behavior in identifying vulnerabilities, raising questions about the comprehensive nature of the safeguards.
A New Framework for Evaluating "Jailbreaks"
In light of the Fable 5 incident, Anthropic is proposing a novel framework for scoring and categorizing AI model "jailbreaks." This framework aims to provide a more structured and objective method for assessing the risks associated with such bypasses. The proposed criteria include the "capability gain" unlocked by the jailbreak, the number of distinct "offensive cybersecurity tasks" for which this capability is applicable, the ease with which the jailbreak can be weaponized, and the difficulty of discovering and obtaining the bypass technique.
Anthropic acknowledges that this framework is currently under development and that determining the precise scoring and weighting of these criteria will require further industry consensus. To support this initiative, the company is establishing a dedicated team to monitor its jailbreak submission channels around the clock. Additionally, Anthropic is launching a new program on the HackerOne platform, inviting security researchers to submit potential vulnerabilities and contribute to the ongoing effort to secure AI models.
Collaborating with Washington and Shaping Industry Standards
Anthropic’s engagement with the U.S. government extends beyond addressing immediate safety concerns. The company is committed to ongoing collaboration with federal agencies, including the Office of the National Cyber Director, the Office of Science and Technology Policy, the Department of the Treasury, and the Department of Commerce. This collaboration is guided by the framework established in the White House’s Executive Order on "Promoting Advanced Artificial Intelligence Innovation and Security."
Anthropic expresses hope that this collaborative approach, combined with its proposed industry framework for assessing AI risks, will lay the groundwork for standardized industry-wide regulations. The company envisions these efforts as a potential template for effective global coordination on managing both the risks and benefits associated with artificial intelligence. Anthropic’s ultimate goal is to see these principles codified into robust regulations that are applied uniformly across all developers of frontier AI models.
As a long-standing advocate for AI safety regulations, Anthropic’s call for industry-wide rules, particularly those that apply to competitors, is consistent with its established position. This proactive stance suggests a strategic effort to shape the future regulatory landscape of AI development and deployment. The reintroduction of Fable 5, therefore, is not just a technical event but also a significant step in Anthropic’s broader mission to foster responsible AI innovation.
