Anthropic–OpenAI feud, Claude Opus 4.5, and FlashAttention 4 shape near‑term backend AI choices
Amid a public Anthropic–OpenAI feud over Pentagon work, Claude model churn and new inference kernels signal fast-moving vendor risk and performance upside for production AI. A Fortune report details escalating tensions between Anthropic and OpenAI tied to U.S. defense work, highlighting how leadership decisions can quickly change enterprise AI risk and procurement posture ([Fortune](https://fortune.com/2026/03/05/anthropic-openai-feud-pentagon-dispute-ai-safety-dilemma-personalities/)). Treat this as an early warning to test exit paths and dual-vendor strategies. Model catalogs like ZenMux list Claude Opus 4.5 with agentic workflow support, long‑horizon reasoning, and stronger prompt‑injection resilience, signaling a step-up for structured, multi-step tasks ([ZenMux](https://zenmux.ai/anthropic)). Evaluate it against your current model on mission‑critical formats and tools. On the infra side, Together AI introduced FlashAttention 4 and highlighted rapid growth, reinforcing that kernel and runtime advances can unlock real serving gains without wholesale re-architecture ([Radical Data Science](https://radicaldatascience.wordpress.com/2026/03/05/together-ai-announces-business-and-product-milestones-at-first-ai-native-conference/)). Consider piloting optimized endpoints where they fit your latency and cost targets.