Anthropic's Claude Mythos 5, codenamed "Capybara," is the first widely recognized 10-trillion-parameter model, currently in limited early-access testing for cybersecurity defense. Using a refined MoE architecture with roughly 1 trillion active parameters per forward pass, it has demonstrated the ability to identify and exploit zero-day vulnerabilities in every major OS and browser during red-team testing. The model was inadvertently revealed through a content management misconfiguration that exposed ~3,000 internal assets in late March.
OpenAI will spend over $20 billion on Cerebras servers over three years — double its earlier $10B commitment — and receive warrants for up to a 10% stake in the chipmaker. OpenAI is also providing Cerebras ~$1B to fund data center development. The deal is central to Cerebras' upcoming IPO, targeting a $35B valuation with a $3B raise next month.
Stanford HAI's annual report shows Anthropic's Claude Opus 4.6 leads ByteDance's best model by only 39 Elo points — a gap that was 17–31 percentage points in 2023. The US still dominates in private investment ($286B vs China's $12B), but China leads in publications, patents, and robot installations. Critically, AI researcher inflow to the US has dropped 89% over seven years.
Google DeepMind released Gemma 4 in four sizes (2B to 31B parameters), all multimodal with text, image, and audio support, under a fully permissive Apache 2.0 license. The 31B model ranks #3 on Arena AI's open model leaderboard. Context windows extend to 256K tokens, with native support for 140+ languages and agentic workflows.
Perplexity rolled out Personal Computer to all Max subscribers ($200/mo) on April 16, an always-on AI assistant that orchestrates across local files, native Mac apps, iMessage, Mail, and Calendar. It can run 24/7 on a Mac mini and be triggered remotely from iPhone via 2FA. This represents a shift toward AI as operating system layer rather than chat interface.
CEO Evan Spiegel told staff that "rapid advancements in artificial intelligence" allow smaller teams to match prior output, with AI now generating over 65% of Snap's new code. The cuts save $500M annually and are expected to establish a path to net-income profitability. Snap stock jumped on the announcement.
PwC's 2026 AI Performance Study finds extreme concentration of AI value — the top quintile of adopters focus on growth (new products, new markets) rather than pure cost-cutting. Companies treating AI as a productivity tool alone are falling behind those using it as a growth engine.
