Claude Mythos’ risks and price
Anthropic’s Claude Mythos is being kept partner‑only because it reportedly shows much stronger offensive and cross‑domain capabilities than earlier models (#). The podcast reported a 72% autonomous exploitation rate on zero‑day tasks versus 14% for Opus, and noted strong performance on biological recovery tasks alongside suggested token pricing of roughly $25 input / $125 output per million tokens (#). Indian fintech firms are already seeking early partner access, signalling immediate interest and concern from financial services about controlled use of the model (x.com).
Anthropic is keeping Claude Mythos off the public market and inside a partner program after tests showed it could find and exploit software flaws far beyond earlier Claude models. (anthropic.com) Anthropic announced Project Glasswing on April 7, 2026, with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia and Palo Alto Networks as launch partners. The company said more than 40 additional organizations that build or maintain critical software infrastructure also received access. (anthropic.com) Anthropic said in its system card that Mythos Preview is its “most capable frontier model” and that the jump in capability led it not to make the model generally available. Bloomberg reported the company has “no plans yet” to release Mythos to the general public while it studies what guardrails are needed. (anthropic.com) (bloomberg.com) In plain terms, the model is being treated less like a normal chatbot and more like a restricted security tool. Anthropic said Glasswing members are using Mythos for defensive work on critical codebases, while the company shares findings across the group. (anthropic.com) Anthropic’s security research site said Mythos can identify and exploit zero-day vulnerabilities, meaning previously unknown software bugs, across every major operating system and major web browser when prompted. SecureWorld reported Anthropic’s tests found a 27-year-old OpenBSD crash bug, a 16-year-old flaw in FFmpeg’s H.264 codec and other issues still under coordinated disclosure. (red.anthropic.com) (secureworld.io) The pricing signal is also unusual. Anthropic’s public API page lists Claude Opus 4.7 at $5 per million input tokens and $25 per million output tokens, while industry trackers and partner reports have pegged Mythos access at about $25 input and $125 output per million tokens. Anthropic has not published Mythos on its public pricing table. (platform.claude.com) (llm-stats.com) (claudefa.st) That puts Mythos at roughly five times Opus 4.7 pricing on both input and output tokens, based on the reported partner rates. Anthropic is also committing up to $100 million in usage credits and $4 million in donations to open-source security groups through Glasswing. (platform.claude.com) (anthropic.com) Bloomberg reported that Indian fintech companies are already seeking early partner access, a sign that banks and payment firms want the model for security work before any wider release. JPMorganChase is already on the launch-partner list, tying the first rollout directly to financial services as well as Big Tech. (x.com) (anthropic.com) Anthropic and its peers are framing the rollout as a race to put stronger AI in defenders’ hands before attackers get similar tools. Newton Cheng, who leads Anthropic’s cyber effort, told Bloomberg the issue is not unique to one company and that Glasswing is meant to give defenders “a head start.” (bloomberg.com) For now, Mythos is being sold as a closed-access model with premium pricing, limited customers and a narrow job: find dangerous bugs before someone else does. That leaves Anthropic testing whether a model powerful enough to automate parts of hacking can stay useful without becoming widely usable. (anthropic.com 1) (anthropic.com 2)
