Claude Code vs. Codex CLI for Terminal Coding

How we testedWe assigned the same three refactors to each tool in the same three repos: rename a domain concept across roughly 40 files in a TypeScript codebase, swap an ORM layer in a Python service, and extract a shared package from a Go monorepo. We graded each attempt on whether the build passed, whether the existing test suite passed, and how many files we had to hand-correct afterward.

Claude Code finished all three refactors with fewer hand-corrections, and the diffs were the ones we would have written ourselves. That matches what independent reviewers report at the benchmark level: <cite index="6-12,6-13">Opus 4.8 hits 88.6% SWE-bench Verified, 69.2% Pro, while GPT-5.5 leads Terminal-Bench at 82.7%</cite>, and on SWE-bench Pro specifically <cite index="6-35">Claude Opus 4.8 leads at 69.2% vs GPT-5.5's 58.6%</cite>. Quality scoring tracks too: <cite index="2-17,2-18,2-19">in blind evaluations where developers rated code without knowing which tool produced it, Claude Code won 67% of comparisons against Codex CLI's 25% (8% were ties). This is the most significant quality gap in the data. Claude Code produces code that human developers consistently judge as cleaner, more idiomatic, and better structured.</cite> Codex wasn't far behind on the small refactors, but on the ORM swap it left more stragglers we had to clean up by hand.

How we testedWe picked four open issues from our test repos (two bug fixes, a small feature, and a dependency upgrade) and assigned each one end-to-end to each tool's agent. We scored whether the agent opened a working PR, how many follow-up prompts we had to give it, and whether the diff was something we would actually merge. We ran each session to completion, including multi-hour runs.

Claude Code held context across the longer-running tickets in a way Codex didn't. Part of that's the model and part is the harness: <cite index="3-23,3-24">Max and Team Premium default to Opus 4.7, and Opus 4.7 exposes a 1M token context window at standard pricing when used (no long-context premium)</cite>, and our tests overlapped with the May 28 Opus 4.8 release that Anthropic shipped mid-cycle. Codex's long-context story is technically comparable: <cite index="3-22">context is 272K by default with a 1.05M-token experimental long-context mode you enable via model_context_window / model_auto_compact_token_limit configuration</cite>, but <cite index="3-28">Codex CLI on GPT-5.4 reaches 1.05M with long-context mode enabled, billed at the 2×/1.5× multiplier when you cross 272K input</cite>, which made us think twice about turning it on for routine work. On the longest run, a multi-step dependency upgrade, Codex lost the thread between sessions. Claude Code didn't.

How we testedWe ran each tool's non-interactive mode against the same three scripted jobs: a unit-test fixer, a small migration generator, and a security review on a deliberately suspicious dependency. We checked the sandbox model, how each tool handled untrusted code, and how cleanly each one dropped into a GitHub Actions pipeline. We also evaluated cross-tool config portability.

Codex is the better fit when isolation and automation matter more than reasoning depth. It's genuinely open-source: <cite index="6-1">Codex CLI is fully open-source under Apache-2.0, Rust-native, with 91,000+ GitHub stars and 800+ releases</cite>, and the sandbox runs at the OS layer: <cite index="29-7,29-8">that sandbox runs on Apple Seatbelt on macOS and Landlock plus seccomp on Linux. Codex CLI is the only major AI coding agent that enforces security at the kernel level, not through application-layer hooks.</cite> For CI, <cite index="29-20,29-21,29-22">the codex exec command runs Codex in non-interactive mode for scripted and CI workflows. It takes a prompt plus stdin, so you can pipe input from another process and pass a separate instruction on the command line. That turns Codex into a tool you can drop into automated pipelines, not just use by hand.</cite> The config file is portable across tools too: <cite index="10-38">AGENTS.md is an open standard governed by the Agentic AI Foundation under the Linux Foundation, adopted by 60,000+ projects</cite>. Claude Code has closed the governance gap with hooks: <cite index="10-15,10-16">Claude Code is better for deep refactoring, code review, and programmable governance through its lifecycle hook system; Codex CLI is better for kernel-level sandboxing and cross-tool portability via AGENTS.md. Claude Code enforces safety at the application layer with more than two dozen hook events you wire up yourself, while Codex enforces safety at the OS kernel layer where the model cannot circumvent restrictions.</cite> For a security-review use case, we picked the kernel.

How we testedWe compared published subscription pricing at every tier and modeled a month of cost for a single heavy user against both the subscription caps and the underlying API rates. We measured token usage per task across our refactors and agentic sessions, and we factored in the prompt-caching, batch, and subscription-versus-API trade-offs each company documents.

The headline prices are close. <cite index="20-8">Claude Code costs $20/month on the Pro plan, $100 or $200/month on Max, or pay-per-token via the Anthropic API.</cite> For OpenAI, <cite index="22-3">ChatGPT Plus, Pro, Business, Edu, and Enterprise plans include Codex</cite>, so if you're already paying for ChatGPT, Codex's CLI has no separate subscription line. Two things tip this round to Codex. First, token efficiency: <cite index="29-44,29-45">token use favors Codex CLI by a wide margin. It burns roughly 3-4x fewer tokens per task than Claude Code, so it costs less per operation at scale.</cite> Second, openness: <cite index="29-37,29-38,29-39">Codex CLI is free and open-source. You pay for the models behind it. OpenAI has moved Codex pricing to a token-based model that lines up with standard API rates.</cite> Claude Code has its own cost levers worth knowing about: <cite index="14-13,14-14,14-15,14-16">Anthropic's Batch API processes requests asynchronously within a 24-hour window in exchange for a flat 50% discount on all input and output tokens. This applies to every Claude model without exception. It's ideal for content generation, data classification, document analysis, and any workload where real-time responses aren't required. The trade-off is simple: if your task can wait, you pay half.</cite> But for everyday interactive work, Codex's lighter token footprint is the larger lever.