OpenAI shipped an April Codex update framed in the coverage as a direct shot at Anthropic's Claude Code, and the framing is fair. The refresh brings three concrete changes. The default model has been GPT-5.4 since March 5, with 1 million tokens of context and improved tool search across large codebases. Codex can now run multiple concurrent agents on a single project, each in an isolated git worktree, which is the pattern Claude Code popularized. And the update ships as part of OpenAI's newly unified desktop superapp that combines ChatGPT, Codex, and the Atlas browser into one environment. The target is clear: match Claude Code's workflow, beat it on cost and speed, and make the OpenAI ecosystem sticky enough that buyers do not bother comparing anymore.
Two numbers matter for anyone actually choosing between the tools. In blind code-quality evaluations earlier in 2026, Claude Code won 67 percent of the time against Codex CLI on equivalent tasks. On the same tasks, Codex used roughly 3x fewer tokens. Both are real, and they are not contradictory. Claude Code's advantage is concentrated in long-context multi-step reasoning where its 1M-token window actually gets used and the agent needs to hold a large plan in working context. Codex's advantage is concentrated in well-bounded tasks that parallelize, where the token-efficiency compounds across the fan-out. The parallel-agents-in-git-worktrees capability is the real new feature to evaluate. It changes the mental model from "one agent, serialized" to "dispatch ten agents, review the ten PRs." That workflow is powerful for certain kinds of work (bug-fix sweeps, dependency bumps, refactor-across-files) and unhelpful for others (single-feature development where coordination overhead swamps the parallelism gain).
The tool competition has matured past feature parity into genuine positioning. Claude Code is the default for complex multi-step reasoning, long-context operation, teams that care about local-execution privacy, and anyone who lives in a terminal. Codex is the default for asynchronous and parallel task delegation, cost-sensitive operations at volume, and teams already embedded in OpenAI's ecosystem. Those are different product bets and both are legitimate. Builders choosing between them should stop looking for a single answer and start routing tasks by shape. The third player worth tracking is Cursor, which is sliding into a role as a neutral multi-model harness: the interface layer that lets teams use Claude for deep reasoning, Codex for parallel dispatch, and a local model for sensitive code, without re-tooling each time. If Cursor gets that layer right, the model choice becomes a configuration decision rather than a platform decision.
For anyone running a coding agent today, three moves follow. First, instrument token cost per task, not just task success rate. Most teams do not know whether they are paying 3x what they need to, because the token number is rarely surfaced in daily workflow. Second, audit which of your coding tasks actually benefit from parallel agents. If your answer is "all of them," you have not thought about coordination overhead honestly; if your answer is "none of them," you are probably wrong about your bug-fix and dependency work. Third, keep your agent harness model-agnostic. Claude Opus 4.7 shipped today, GPT-5.4 is six weeks old, and the next iteration is always a few weeks away. Whichever tool wins in April 2026 will not necessarily be the one you want by Q3, and rewriting your agent loops every quarter is not a sustainable posture.