Briefing — May 8, 2026

Today's headlines: • OpenAI adds realtime voice models and starts testing ChatGPT ads — OpenAI shipped more capable speech-native API models while also opening an ad experiment that signals a new monetization phase for consumer AI. • Anthropic’s Mythos shows real traction in AI-assisted security research — Mozilla says Mythos surfaced 271 Firefox vulnerabilities with almost no false positives, making one of the clearest enterprise-grade cases for model-driven bug finding. • Moonshot AI raises $2 billion as China’s open-model market heats up — The round reportedly values Moonshot at $20 billion with ARR above $200 million, underscoring that Chinese AI demand is broadening beyond hyperscalers. • Google shuts down Project Mariner — Google killed its experimental web agent product and says the technology will be folded elsewhere, a reminder that browser agents still struggle with product fit and reliability. • Gemma 4 gets a major inference speedup via future-token prediction — Google says Gemma 4 can run up to 3x faster without quality loss by predicting ahead, a notable sign that speculative-style decoding tricks are moving into mainstream open models. • Anthropic expands Claude Code limits and Managed Agents capabilities — Higher usage caps and 'dreaming' behavior for managed agents point to a more persistent, tool-using coding workflow as vendors race to own the dev loop. • Infrastructure pressure keeps rising from fabs to power grids — TSMC is leaning harder into wind power as AI chip demand strains Taiwan’s energy system, while SpaceX’s reported chip-manufacturing ambitions show the stack keeps integrating downward. • Spotify bets on AI-generated personal audio — New tooling to import AI-generated podcasts into Spotify suggests a viable consumer surface for agent-produced summaries, research digests, and personalized media.