The week every major AI lab redrew its battle lines
OpenAI bets the farm on coding agents, Anthropic weaponizes its strongest model for defense, and Meta breaks its own open-source playbook. The vendor landscape just shifted beneath your feet.
This was the week the AI vendor landscape stopped being about who has the best chatbot and started being about who owns the developer workflow. Three moves, from three different companies, each with very different philosophies — and together they tell you exactly where this industry is headed.
OpenAI went all-in on Codex, rebranding it from a research curiosity into its flagship coding agent. It now lives inside ChatGPT, VS Code, the terminal, and a dedicated web app. The company simultaneously restructured its pricing: a new $100/month Pro tier sits between Plus and the existing $200 tier, explicitly designed around Codex usage. The models powering it — gpt-5.4 for heavy lifting, gpt-5.3-codex-spark for real-time iteration — suggest OpenAI sees coding as the monetization wedge for its entire platform. When you ship a coding agent across every surface your users touch, you're not selling a tool. You're selling a workflow.
Anthropic took a sharply different path. Project Glasswing introduced Claude Mythos Preview, a model that outperforms Opus 4.6 but will never be made generally available. Instead, Anthropic is deploying it exclusively for cybersecurity — partnering with AWS, Apple, Google, Microsoft, NVIDIA, and others to find vulnerabilities in critical infrastructure. The model has already discovered thousands of high-severity flaws across every major operating system and browser. Anthropic committed up to $100M in usage credits for security partners. This is a deliberate strategic choice: demonstrate frontier capability by pointing it at defense rather than productivity, and build trust with exactly the institutions that will regulate AI.
Meanwhile, Meta released Llama 5 on April 8 while simultaneously launching Muse Spark — the first model from its Superintelligence Labs — as closed source. You cannot download the weights. For a company that built its entire AI brand on open-source accessibility, this is a significant break. As The Register noted, "this is no Llama." Meta appears to be splitting its strategy: Llama stays open for community adoption and ecosystem lock-in, while frontier capabilities get the proprietary treatment.
Google kept its head down and shipped. Gemma 4 crossed 10 million downloads in its first week, with the full Gemma family now past 500 million. The models run locally, support agentic workflows, and are Apache 2.0 licensed. Google also quietly turned AI Studio into a full-stack vibe coding environment. No press events, no pricing drama — just distribution.
For builders, the implication is clear: vendor lock-in is no longer about which model you call. It's about which agent framework owns your git workflow, your IDE, and your CI pipeline. For buyers evaluating enterprise AI, watch the Anthropic security play — organizations that get early Glasswing access will have a meaningful defensive advantage. For everyone else, the Sora shutdown and OpenAI's new ad-supported free tier are reminders that consumer AI products remain economically fragile.
The bottom line: The AI vendor war has moved from model benchmarks to workflow control — pick your tools based on where the agent lives, not just how it scores.