Anthropic has launched Claude Sonnet 5 and restored Fable 5 and Mythos 5 after an 18-day interruption triggered by US export controls. A new classifier blocks the jailbreak at the centre of the dispute, allowing the frontier models to return across direct and partner platforms.
The episode connects model release, cyber policy and global availability in an unusually visible way. Sonnet 5 now arrives as the commercial workhorse while Anthropic and regulators attempt to define what constitutes a serious safety breach.
A targeted classifier reopened global access
The shutdown followed reports that Fable 5 could be prompted to identify vulnerabilities and provide exploit code. Anthropic says similar behaviour appeared in older models from several providers, complicating the argument that one frontier release created a unique capability.
Engineers trained a classifier against the bypass, reporting more than 99% prevention in internal trials. Ambiguous requests may be routed to Opus 4.8, preserving service at lower capability.
The wider safety margin creates false positives for benign debugging. Developers need visibility when routing occurs and a way to challenge blocks, particularly in legitimate security work.
Restoration does not erase the availability lesson. Anthropic’s inability to filter users by nationality forced a global blackout, showing that regulatory controls can become operational outages.
Sonnet 5 targets agent workflows at a lower price
Anthropic reports improved coding and terminal benchmarks over Sonnet 4.6 while retaining base pricing of $3 per million input tokens and $15 per million output tokens, with introductory discounts through August.
At Rakuten, the model processed difficult pull requests, ran tests and presented changes for human approval. Zapier used it for multi-stage administrative work, and Zed reported autonomous bug reproduction, repair and verification.
Those examples show the importance of completing a chain rather than generating one good step. Enterprises should still measure correction, tool failure and total token use across the complete workflow.
Permission boundaries must sit outside the model. Updating account tiers and sending announcements combine reversible analysis with consequential action; each deserves separate authority.
Enterprises should also test the effect of the new classifier on their own engineering workload. A security improvement that repeatedly blocks legitimate debugging can push staff toward weaker models or unsafe workarounds. Logging route changes, appeal outcomes and task completion gives Anthropic evidence for tuning without asking customers to accept an invisible productivity penalty.
Shared security metrics could reduce regulatory uncertainty
Anthropic, Amazon, Microsoft and Google are developing a framework that scores breaches by capability gain, breadth, weaponisation effort and discoverability. A common language could help providers distinguish a nuisance bypass from an urgent public risk.
The framework needs independent participation and transparent thresholds. Providers have commercial incentives to minimise disruption, while governments may overreact without comparable evidence.
Anthropic has also added continuous monitoring, a HackerOne programme and pre-release access for federal evaluators. External testing can strengthen safety if it does not become a political veto lacking clear standards.
The Sonnet 5 system card reports lower problematic behaviour than its predecessor and no successful exploit generation in a Firefox assessment, though partial progress increased modestly. Such evaluations should be repeated as tools and scaffolding change.
This release is therefore two stories: a more economical agent model and a repaired relationship between deployment and oversight.
Customers will value Sonnet 5 if it completes long tasks reliably. They will trust Anthropic’s frontier programme only if access, safety and appeals become predictable. The 18-day pause showed that model governance is no longer an internal laboratory process; it can determine whether global businesses wake up with the capability they built around still available.
They also need advance notice when classifiers or routes materially change production behaviour.
