Anthropic is set to release its most potent AI model, Claude Mythos, to customers in the coming weeks, the company confirmed via Reuters. The model, which possesses expert-level hacking abilities—including the autonomous discovery and exploitation of zero-day vulnerabilities—has been under strict internal controls. Only a select group of tech heavyweights such as Apple, Microsoft, Amazon, Google, and CrowdStrike had access through Project Glasswing, a defensive testing program. With additional safeguards in place, Anthropic now plans a gradual rollout to enterprises.
Mythos: A Dual-Edged Sword in Cybersecurity
First teased in April, Mythos is Anthropic's most advanced frontier model, showing major leaps in general capabilities. Its standout feature is the ability to autonomously find and exploit zero-day vulnerabilities in operating systems, browsers, and software, raising global concerns about weaponization. To mitigate risks, access was initially limited to defensive applications under the Glasswing project. Now, as security measures mature, Anthropic is ready to expand access to vetted clients.
Opus 4.8 Goes Live: Faster, Cheaper, and More Honest
Alongside the Mythos announcement, Anthropic today launched Claude Opus 4.8, an upgrade priced identically to its predecessor ($5/$25 per million input/output tokens). Key enhancements include: a 2.5x faster Fast Mode at one-third the cost of the previous fast mode, ideal for high-throughput tasks; improved performance in SWE-Bench and Terminal-Bench; support for dynamic workflows with hundreds of parallel sub-agents; and a major honesty upgrade—early testers report the model now flags uncertainty rather than confidently producing false conclusions, reducing hallucinations. Alignment reportedly approaches Mythos preview levels.
With Opus 4.8 rolling out and Mythos set for imminent release, Anthropic is executing a dual-track strategy of safety upgrades and extreme performance to dominate the enterprise AI market.