Three Quarters of Google Cloud Runs AI, and Fear Still Leads the Tape

Claude Mythos Takes the Lead

Anthropic's latest model has seized the top spot on frontier benchmarks, a position that shifts competitive dynamics across the industry. With Google, OpenAI, and Anthropic trading the lead on a near-monthly cadence, the performance gap between top models is narrowing even as capability ceilings rise. The real question is whether benchmark supremacy translates into enterprise revenue or remains a prestige metric.

South Korea's Fair Use Framework

Seoul has published formal guidance on fair use for AI training data, staking out a middle path between the EU's restrictive copyright regime and the largely unregulated U.S. approach. The guide gives Korean AI companies clearer legal footing to train on publicly available content while preserving opt-out mechanisms for rights holders. This positions South Korea as a potential regulatory model for other mid-sized tech economies navigating the same tension.

Local Inference Breaks Speed Barriers

On-device AI models are now hitting 6,000 tokens per second, a throughput level that makes local inference competitive with cloud APIs for many production tasks. This milestone undermines the assumption that frontier capability requires centralized compute, opening the door for privacy-sensitive and latency-critical applications to run entirely on consumer hardware. For cloud providers like Google, whose 75% AI adoption figure headlines today, the long-term threat is that the most cost-conscious workloads migrate off their platforms entirely.