Why developers are quietly switching from Claude Code to Codex

OpenAI did not ship one product update in April. It shipped five — and stitched them into a single product. Between April 16 and May 7, 2026, Codex went from a coding agent that lived inside a window into a full developer workstation that controls a Mac, runs in a Chrome tab, remembers your preferences, talks to 90+ third-party tools, and schedules its own follow-up work. The number that matters: 3 million developers are using it every week.

The single biggest move sits in the April 16 announcement, "Codex for (almost) everything." That was the day OpenAI flipped Codex from "AI that writes code" into "AI that uses the apps you write code with." Background computer use lets a sandboxed cursor click and type across any Mac application while you keep working in your own apps. Multiple agents can run in parallel without interfering with each other. The follow-on, three weeks later on May 7, extended the same logic into the browser via a Chrome plugin that handles DevTools work and cross-tab context.

Why this matters: the bet is that the unit of "coding work" is no longer "edit a file in an IDE." It is "complete a task that touches a Mac, a browser, a Linear ticket, a Figma file, a deploy log, and a Slack channel." Codex is now structured to do all of that in one session. Claude Code, Cursor, and the other named players still mostly assume the developer is the one moving between those surfaces. That is the gap OpenAI is trying to make permanent.

The numbers nobody is repeating: 3 million weekly active developers. 90+ plugins added in a single drop, including docs, project management, code review, creative tooling, and deployments. 1,000 tokens per second on Cerebras hardware versus roughly 200 for Claude Code. 2 to 3× faster iteration cycles reported in independent reviews for frontend, testing, and design handoff workflows. And a single $20 ChatGPT Plus tier that grants 30 to 150 Codex sessions per 5-hour window — the same tier that hosts everything else OpenAI sells.

The backstory: Codex desktop launched in February 2026 as a relatively conservative agent that could read a repo and write code. Two months later it has Computer Use, an in-app browser built on OpenAI's Atlas technology, gpt-image-1.5 for inline image generation, persistent memory of past actions, and a scheduling layer that lets agents resume work days or weeks later. The April release also added Intel Mac support — previous versions had been Apple Silicon only — which quietly broadened the addressable developer base by tens of millions of machines.

The countervailing angle Anthropic users will want to read twice: Claude Code is still better at the things developers say they care about. Claude Opus 4.6 leads SWE-bench Verified at 80.8%. Claude's 1M-token context (in beta) dwarfs Codex's 400K. VS Code extension ratings show Claude at 4.0 out of 5 versus Codex at 3.4. Blind code reviews of identical prompts rate Claude's output as cleaner, more idiomatic, and more consistent with existing project structure. One Reddit comment cited in multiple comparisons captured the trade-off: "Claude follows the plan. Codex goes off-plan most of the time." And yet 65% of those same surveyed developers still use Codex daily — because the speed, the terminal accuracy (77.3% on Terminal-Bench 2.0), the built-in computer control, and the plugin breadth outweigh the output-quality gap for most daily work.

What to watch over the next 1-3 months: (1) Anthropic's response to background computer use — Claude has had Computer Use as a research preview since 2024 but has not shipped a parallel-agent desktop app at OpenAI's scale. (2) Whether Codex's 90+ plugin marketplace becomes a moat or a liability — third-party prompt-injection surface area is now enormous. (3) GPT-5.3-Codex's lead on Terminal-Bench 2.0 versus the next Claude release. (4) Whether Cursor and Replit pivot to integrate Codex itself, follow Anthropic, or build their own.

The takeaway: a coding agent that controls your computer is no longer the future. It shipped on April 16. The race that matters now is not who writes the cleanest code — it is who controls the most surfaces while writing it.