Anthropic, OpenAI models alarm Washington

Anthropic and OpenAI are now colliding in Washington on two fronts: model capability and model economics. Researchers who tested Anthropic’s Mythos and OpenAI’s GPT-5.5 told Politico that the systems’ hacking performance has changed how security officials are assessing frontier AI risk. (yahoo.com) OpenAI, meanwhile, has advertised a Preparedness role paying up to $445,000 to study “recursive self-improvement,” a posting that put unusually explicit risk language into public view. (thebridgechronicle.com) At the same time, DeepSeek said it would make a 75% price cut on its V4-Pro model permanent, extending a discount that further lowers the cost of access for developers. (thenextweb.com) ### Why are Washington officials focused on Mythos and GPT-5.5? Lee Klarich, chief product and technology officer at Palo Alto Networks, told Politico that Anthropic’s Mythos was a “game-changer” when he first tested it. (yahoo.com) Politico reported that researchers with access to Mythos and OpenAI’s GPT-5.5 said the models’ hacking capabilities have become central to U.S. security assessments. May 7 added a second datapoint. (thebridgechronicle.com) Politico reported earlier this month that OpenAI introduced GPT-5.5-Cyber, a restricted model for vetted cybersecurity professionals, two weeks after the broader GPT-5.5 release and after Anthropic had limited Mythos access to a small group of testers. (thenextweb.com) ### What makes the OpenAI job listing notable? OpenAI’s posting was for a safety researcher on its Preparedness team, with compensation reaching $445,000 a year, according to reports published May 24. The role focuses on “recursive self-improvement,” described in coverage of the listing as the possibility that AI systems could accelerate their own capabilities by helping train or design successor systems. (yahoo.com) The wording drew notice because it framed the work as preparation for a risk scenario rather than a product launch. Reports on the listing said OpenAI was seeking people to support preparations for self-improving systems and highlighted language saying the work required reasoning about problems that “might exist in the future.” (politico.com) ### Where does DeepSeek fit into this story? DeepSeek made a pricing move on May 24 that affects the market even if it does not change frontier capability. The company said a 75% discount on its V4-Pro model would become permanent, keeping output-token pricing at $0.87 per million, according to reports on the announcement. (thebridgechronicle.com) Those prices matter because they change who can afford large-scale testing, coding workloads and evaluations. The Next Web reported that the new baseline leaves V4-Pro at a fraction of the cost of some U.S. rivals, extending a price war that has already pushed developers to compare model quality against operating cost, not just benchmark scores. (thebridgechronicle.com) ### Are capability fears and price cuts happening at the same time? May 2026 suggests yes. Anthropic and OpenAI are restricting some advanced cyber-oriented access even as companies race to ship stronger systems, and OpenAI’s own hiring shows the company is dedicating staff to explicit long-range risk work. (thenextweb.com) The result is a market where the most advanced systems are being treated as security-sensitive while cheaper access to powerful alternatives keeps expanding. That combination — higher concern at the frontier and lower prices below it — is visible in the Mythos and GPT-5.5 reporting and in DeepSeek’s new pricing. (thenextweb.com) ### What comes next? OpenAI’s next visible step is hiring for the Preparedness role it has already posted, while access to GPT-5.5-Cyber remains limited to vetted cybersecurity users, according to published reports. Anthropic’s Mythos access is also still restricted to selected testers. (politico.com) DeepSeek’s next milestone is simpler: the discounted V4-Pro pricing is now the standing rate rather than a temporary promotion. That leaves Washington watching one set of companies for cyber capability and the broader market watching another for how low model access costs can go. (thenextweb.com) (politico.com) (yahoo.com)