Codex Becomes a Platform, Anthropic Becomes a Weapon, and Vendor Lines Blur

OpenAI is repositioning Codex as the operating layer for knowledge work, while Anthropic's models are quietly being pulled from public release for being too capable at finding zero-days. The vendor map is getting weirder.

Three things shifted this week that change how working teams should think about vendor selection. OpenAI stopped pretending Codex is a coding tool. Anthropic published a study of 832 malicious accounts that doubles as a warning about its own models. And Apple — the company that built its identity on silicon independence — is reportedly routing its Gemini-powered Siri through Nvidia chips. The neat categories of 2025 are dissolving.

The Codex repositioning is the most concrete. According to OpenAI, Codex now ships with 62 app integrations and 110 role-specific skills covering sales, data analytics, product design, and creative production. A separate Sites feature, rolling to Business and Enterprise plans as of June 2026, lets Codex generate interactive web apps from a brief. As 9to5Mac reports, Codex is moving inside the ChatGPT app everywhere over the next few weeks. The framing has shifted from autocomplete to general-purpose knowledge work — a direct play for the seat that Cursor, Claude Code, and Copilot have been fighting over at the developer tier.

Meanwhile Anthropic is having a stranger week. The company published research mapping 832 malicious accounts against the security community's existing taxonomy of threat-actor techniques — useful defensive work. But Simon Willison noted on X that a recent Claude variant was reportedly held back from general availability because it was finding zero-day vulnerabilities at an uncomfortable rate. Capability is now a release-gating concern in a way it wasn't 18 months ago. Separately, Anthropic endorsed a new White House Executive Order on AI, continuing its pattern of cultivating policy proximity that OpenAI has historically owned.

What this means for builders

If you're a developer choosing tools, the pricing comparison from Developers Digest (published 3 days ago) still puts Cursor and Claude Code ahead of Copilot on agent capability at the $10–20/month tier. But the Codex plugin push means the comparison set is about to expand beyond the IDE. If you're a buyer, watch the Apple-Nvidia-Gemini triangle: when the most vertically integrated company in tech outsources inference silicon, the assumption that hyperscalers will own the AI compute stack gets shakier.

The bottom line: Re-evaluate your vendor shortlist quarterly now — the categories that defined 2025 procurement decisions are collapsing into each other faster than annual contracts can keep up.

Sources