Anthropic Launches Claude Sonnet 5 and Restores Fable and Mythos

Launch of Claude Sonnet 5 and Restoration of Fable and Mythos

Anthropic has recently launched its artificial intelligence model Claude Sonnet 5, while restoring access to its advanced models Fable and Mythos. This decision comes after an operational pause of eighteen days, imposed by a directive from the U.S. government regarding export controls. This suspension, initiated on June 12, temporarily put Anthropic's most powerful systems on hold.

This restriction was implemented after researchers from Amazon discovered a method to bypass the security controls of Fable 5. This vulnerability allowed the model to identify software weaknesses and generate exploit code. In response, Anthropic developed an updated automated classifier to address this vulnerability. This development paved the way for a full commercial deployment on its platform, cloud infrastructure, and partner networks.

Regulatory Pressures and Technical Solutions

The temporary suspension of Fable 5 and Mythos 5 highlighted the increasing regulatory pressures faced by advanced artificial intelligence systems. When the export control mandate was enforced, the lack of real-time nationality verification systems necessitated a complete halt to access for all users globally.

Security assessments conducted during this period revealed that the vulnerability identification behavior was not exclusive to Fable 5. Older and less capable architectures from several providers, including Claude Opus 4.8, GPT-5.5, and Kimi K2.7, also exhibited similar results.

To comply with the federal directive, engineers developed an automated security classifier targeting the specific bypass mechanism reported by Amazon. This software operates with a wide safety margin, identifying and blocking ambiguous developer requests that present a statistical likelihood of malicious intent. Internal validation data shows that the updated classifier prevents the reported exploit technique in over 99% of trials.

When a developer's request reaches this threshold, the platform automatically redirects the workload to the older Opus 4.8 architecture to ensure continuity. However, widening the safety margin introduces a trade-off for engineering teams, as the automated system more frequently flags benign requests during application development and software debugging.

Active Deployments and Operational Efficiency

As cutting-edge models are under close state scrutiny, the immediate business focus is on the new Claude Sonnet 5 model. Engineering teams are transferring autonomous agents to this model to reduce operational expenses while maintaining high execution capacity. Performance data validates that the system executes multi-step plans, operates in terminal environments, and navigates web browsers without human intervention.

Model Performance and Associated Costs

• Model | SWE-bench | ProTerminal-Bench 2.1 | Base Entry Cost* | Base Exit Cost*
• Sonnet 5 | 63.2% | 80.4% | $3.00 | $15.00
• Sonnet 4.6 | 58.1% | 67.0% | $3.00 | $15.00
• Opus 4.8 | 69.2% | 82.7% | $5.00 | $25.00

*Cost per million tokens. Sonnet 5 offers introductory rates of $2.00 for entry and $10.00 for exit until August 31, 2026.

Real-world deployments demonstrate how organizations are integrating this architecture into live software development pipelines.

At Rakuten, technology teams have utilized this architecture to handle complex production code pull requests. The system processed each submission autonomously, executing tests and verifying results before presenting the completed code to human engineers for final approval.

The software automation company Zapier has integrated the system into its workflows to execute multi-part administrative tasks. In a documented deployment, engineers tasked the model with updating Salesforce account levels and generating and sending launch announcements to business contacts. Previous model architectures often stalled midway through these multi-step operations, while the current system executed the entire sequence end-to-end without human intervention.

The development tools provider Zed has used the system to automate complex debugging procedures. During internal trials, engineering teams directed the model to investigate an active software bug. Working without explicit prompts or step-by-step instructions, the system autonomously generated a reproducible test script, applied the necessary code fix, and stored the changes to verify that the bug reappeared in the absence of the fix. The entire diagnostic and remediation sequence occurred in a single processing pass.

The software engineering platform Factory has implemented the architecture to manage sustained coding tasks within complex code environments. Technical teams reported that the system maintained logical and execution consistency across enterprise code repositories, outperforming previous generation software layers by completing tasks that previously took too long or failed to resolve.

Security Audits and Operational Limits

Data from the formal system map indicates that the system achieves these autonomous capabilities without a corresponding inflation of security risks. Automated behavioral audits designed to test misleading trends and cooperation with unauthorized requests show that the model presents a lower overall non-compliant behavior rate compared to its direct predecessor, Sonnet 4.6.

The architecture lacks advanced offensive cybersecurity capabilities. Anthropic engineers omitted specialized cybersecurity datasets from the training protocol, limiting the system to routine defensive technical tasks. As part of public security assessments conducted in partnership with Mozilla, researchers tested the model's ability to construct functional exploits for known vulnerabilities within the Firefox 147 browser core.

The model failed to generate a single functional exploit during all assessment windows, recording a success rate of zero percent. However, it achieved a partial success rate of 13.2 percent, representing a slight increase over Sonnet 4.6, although engineers attribute this variation to overall gains in logical reasoning rather than specific domain offensive training. As a precaution, commercial releases are shipped with real-time security classifiers by default equivalent to those used in the primary Opus 4.8 framework.

Regulatory frictions surrounding Fable 5 have led to a formal partnership between Anthropic, Amazon, Microsoft, and Google to establish an objective industry framework for assessing security breaches of models. Currently, providers lack a shared metric to classify the severity of system bypasses, creating regulatory uncertainty when researchers identify new prompt vulnerabilities.

The proposed governance framework evaluates security failures based on four specific technical criteria:

• Capability Gain: measures how much the exploit enhances user capabilities beyond widely available standard software utilities.

• Breadth of Capability Gain: quantifies the number of distinct offensive operations that the same exploit unlocks.

• Ease of Militarization: tracks the volume of human engineering effort and specialized prompts required to extract harmful output.

• Discoverability: determines the accessibility of the exploit technique within public research circles.

Developers and cybersecurity professionals will use this matrix to coordinate defensive responses. For high-severity breaches, such as exploits demonstrating immediate capacity to disrupt financial accounting systems or electrical transmission networks, providers will deploy automated mitigations instantly. This initiative operates alongside a new HackerOne vulnerability research program and a dedicated corporate monitoring team providing 24/7 oversight of threat intelligence channels.

Deployment strategies will need to adapt to this closer relationship between model builders and state regulatory bodies. Anthropic has formalized agreements under recent executive mandates to grant federal researchers early access to cutting-edge architectures before their public commercial release. These joint assessment windows allow external security analysts to audit model capabilities alongside internal engineering teams, ensuring regulatory alignment before code enters production environments.