Codex

OpenAI's Codex went from roughly 1 million to roughly 8 million active users in five months, folded into ChatGPT Desktop alongside ChatGPT Work, and became a frontal competitor to Claude Code. The adoption curve is the least interesting thing about it.

What keeps bringing it back is its role as the worked example of the model-plus-harness pattern. In February 2026 OpenAI described an internal beta of about a million lines of code built by three engineers (later seven) with zero lines written by hand: Codex agents opened, evaluated and merged roughly 1,500 pull requests, about 3.5 per engineer per day, the first commit coming from Codex CLI running GPT-5. OpenAI separately reported that Codex CLI had merged more than two million public pull requests on GitHub. Vivek Trivedy of LangChain notes the coupling this creates: Codex and Claude Code are post-trained with their harness, so changing a tool's logic can degrade performance.

Eric S. Raymond gives the practitioner's view. Across 63 projects in C, Go, Rust, Python and shell, the old failure of a model drifting near its context limit was replaced by a red warning prompting a session flush; his routine is regression tests, then asking Codex to audit the code for bugs and improvements.

The unresolved part is coexistence. Block had access to goose, Claude Code and Codex, each with different interfaces, configuration systems and ways of managing context, and built Berd rather than another agent.