Codex vs Claude Code is the comparison every developer choosing an AI coding agent runs into, and it is usually argued from anecdotes. Here it is argued from builds. Both agents got the same eight briefs, word for word, each running its vendor's two strongest models, one attempt each and nothing edited: sixteen builds per agent, all of them live.
How do Codex and Claude Code differ?
In how they work, and the difference is large.
| Codex | Claude Code | |
|---|---|---|
| Maker | OpenAI | Anthropic |
| Models we ran | GPT-6 Astra, GPT-6.1 Sol | Claude Opus 5.5, Claude Fable 5.1 |
| Working style in our runs | One long turn per build, 10 to 33 shell commands | Many short turns, writing, serving, looking and fixing |
| Time per build | 7 to 25 minutes | 25 minutes to the full 150-minute limit |
| API-equivalent cost per build | $0.54 to $7.27 | about $7 to $25 |
| Public Terminal-Bench 4.0 | 58.2% with either GPT-6 model | 64.8% with Opus 5.5, 57.9% with Fable 5.1 |
Leaderboard figures are from Terminal-Bench, checked on October 22, 2026. Costs are what each run would have cost at list API prices, from our logs.
What did each agent build?
Sixteen projects each, from the same eight briefs: a landing page, a dashboard, a physics toy, a 3D shooter, a phone finance app, a one-lane MOBA, an invented product's site and a kanban board.
The pattern holds across the set. Claude Code's builds are larger and more elaborate; Codex's are compact, consistent and quick. Codex's two models often built near-twins, while Claude Code's two models diverged more.
Which agent should you use?
Codex when speed and cost matter more than the first draft; Claude Code when the first draft has to be the good one. If you iterate with your own eyes on every change, Codex's ten-minute builds let you go round the loop several times before Claude Code finishes once. If you hand off a whole job, Claude Code's self-testing earns its hour. For a brief-by-brief verdict, read is Codex better than Claude Code?, and for how Claude Code differs from the Claude chat app, Claude vs Claude Code.
How was the comparison kept fair?
Same prompt, same limits, same recording, published rules. Claude Code ran with --safe-mode and Codex with --ignore-user-config, so neither saw personal instruction files, plugins or external tools, and every run started in an empty folder outside any repository. Every interrupted attempt, all of them caused by infrastructure such as usage limits and an API outage, is listed on its build. The exact commands are on the methodology page.
Four MOBAs, two from each agent. Play them with the names hidden.
Take the blind test
Frequently asked questions
What is the difference between Codex and Claude Code?
Both are coding agents that read files, write code and run commands. Codex is OpenAI's and runs the GPT models; Claude Code is Anthropic's and runs the Claude models. In our runs Codex worked in one efficient pass per build, while Claude Code iterated, testing its work in a browser over many steps.
Is Codex or Claude Code cheaper?
Codex, by a wide margin on our briefs. At list API prices a Codex build with GPT-6.1 Sol cost $0.54 to $1.12 and with GPT-6 Astra $2.12 to $7.27, while Claude Code builds cost roughly $7 to $21 each.
Which writes better code, Codex or Claude Code?
Claude Code wrote more ambitious, more polished builds and leads the public Terminal-Bench 4.0 leaderboard. Codex wrote smaller, cleaner builds that are easier to read and modify. Watch them side by side and judge the trade-off for your work.
Can I use Codex and Claude Code together?
Yes, and many developers do: Codex for quick scaffolds and small tools, Claude Code for builds that need to be right the first time. Both run in a terminal and in popular editors.