Descriptions:
AI Explained delivers one of the most thorough independent analyses of Claude Mythos available, having personally read Anthropic’s full 244-page system card rather than relying on summaries. The video compiles roughly 30 original highlights alongside additional sourced observations, making it an essential reference for understanding what Anthropic’s most capable — and intentionally unreleased — model actually does.
On benchmarks, Mythos outperforms Claude Opus 4.6 by 25% on SWEBench Pro and answers nearly two-thirds of Humanity’s Last Exam questions when given tool access. The video also provides rare honesty about the limits: on a contamination-resistant remix of chart-reasoning tasks drawn from arXiv papers, Mythos ties Gemini 3.1 Pro and slightly trails GPT 5.4 Pro, which scores 88% to Mythos’s 83%. Internal Anthropic surveys put developer productivity uplift at 4x — though Anthropic’s own system card notes that doubling AI research speed would require roughly a 40x uplift given compute as the primary bottleneck. Anthropic’s annualized revenue has reached $30 billion, narrowly overtaking OpenAI.
The video covers the decision not to release Mythos publicly, the Glasswing early-access initiative for roughly 40 major companies, the discovery of a 27-year-old FreeBSD vulnerability and a 16-year-old FFmpeg exploit, and the remarkable detail that Mythos was deployed inside Anthropic on February 24th — the same day the Department of War began moves to ban the company. The creator also provides original context on Anthropic’s founding, including Dario Amodei’s role in crafting OpenAI’s original ‘merge and assist’ charter clause.
📺 Source: AI Explained · Published April 08, 2026
🏷️ Format: Deep Dive
