Anthropic Lets Claude Mythos Spread Its Glasswings
THE SO WHAT
Anthropic widening access to Claude Mythos underlines a new category — models explicitly tuned for exploring values, meaning, and long-horizon questions, but still gated as “dangerous, but not too dangerous.” If your org is using frontier models for strategic or ethical deliberation, you now need an internal policy on when that’s acceptable and how those outputs are audited.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIMicrosoft and Mayo Clinic partner for an AI model trained on Mayo's medical data, with plans to build an AI healthcare assistant and AI tools for clinicians (Clare Duffy/CNN)
Healthcare AI is consolidating around closed, institution-specific models trained on proprietary data — the moat is longitudinal clinical data access, not model architecture. If you run a hospital or payer and don’t have a data-partnered model program in motion this year, you’re training your clinicians on someone else’s stack and someone else’s incentives.
Applied AIAn interview with Microsoft AI CEO Mustafa Suleyman about catching up with "what was state of the art just a few months ago", refusing to use distillation, more (Reed Albergotti/Semafor)
Refusing to use distillation is a statement that quality — not unit cost — is the primary competitive axis at the frontier. If you’re betting your roadmap on small, cheap models, assume your users will compare them against undistilled frontier behavior and notice the gap.
Mercor's CEO says it now spends more on AI tokens than employee salaries
When a startup spends more on AI tokens than payroll and predicts enterprises will spend more on compute than headcount in five years, it’s saying the labor line item is becoming capex to the cloud. If your P&L doesn’t explicitly model “AI spend as workforce,” you’re flying blind on both margin and hiring strategy.
Applied AIUber caps employee AI spending after blowing through budget in four months
Encouraging “use AI everywhere” without a spend architecture just turned into a budget fire drill — usage-based AI is cloud overrun 2.0. If you don’t have per-team AI cost guards and ROI tracking in place, your 2026 opex plan is fiction.