Anthropic launches Claude Sonnet 5, restores Fable and Mythos

Anthropic launched Claude Sonnet 5 and restored access to its Fable and Mythos frontier models after completing an 18-day federal export-control review. The pause began after a June 12 US government directive that required operators to suspend access to their highest-capability systems.

The suspension followed Amazon researchers reporting a prompting technique that caused Fable 5 to identify software vulnerabilities and generate exploitation code. Regulators required a temporary global blackout while providers lacked real-time nationality verification tools to limit access.

During the pause, Anthropic ran targeted security audits and developed an automated safety classifier to block the specific bypass described by the researchers. Internal validation shows the classifier prevented the reported technique in more than 99% of trials. When a prompt crosses the classifier’s conservative threshold, the platform routes the workload to the older Opus 4.8 model to maintain service continuity.

The company acknowledged the expanded safety margin increases the frequency of benign developer requests being flagged, creating a trade-off between tighter controls and occasional disruptions to debugging and development workflows.

Testing during the review found the vulnerability pattern was not unique. Older or differently architected models from multiple vendors, including Opus 4.8, GPT-5.5 and Kimi K2.7, produced similar outputs under the same prompting technique. Anthropic helped establish a joint industry framework with Amazon, Microsoft and Google to standardise how model security breakdowns are scored across capability gain, breadth, ease of weaponisation and discoverability.

Anthropic announced a commercial rollout of Claude Sonnet 5. The company says Sonnet 5 delivers improved multi-step planning and autonomous task execution, with stronger performance on software engineering and terminal benchmarks than its predecessor. Introductory pricing is $2.00 per million input tokens and $10.00 per million output tokens through Aug. 31, 2026; standard rates are higher.

Enterprises have integrated Sonnet 5 into production pipelines. Rakuten used the model to review and test complex code pull requests, executing tests and verifying results before human review. Zapier used it to complete multi-step administrative tasks, including updating CRM account tiers and sending launch communications. Development toolmaker Zed used the model to generate reproducing tests, apply fixes and verify outcomes. Factory applied Sonnet 5 to long-running coding tasks across large codebases and reported improved consistency compared with earlier generations.

The company excluded specialised exploit datasets from Sonnet 5’s training and ships commercial versions with default real-time safety classifiers similar to Opus 4.8. In a public security assessment run with an external partner, Sonnet 5 failed to produce a working exploit for known Firefox 147 vulnerabilities, recording a 0% success rate for functional exploits and a 13.2% partial success rate that the company attributes to general reasoning improvements.

To strengthen defensive capacity, Anthropic and partners plan a HackerOne vulnerability research programme and a corporate monitoring team offering 24-hour threat intelligence oversight. Anthropic has formalised arrangements to give federal researchers early access to frontier models prior to public release, allowing external audits to run alongside internal testing before models are commercialised.