Is Codex better than Claude Code? On speed and cost, clearly yes. On depth and polish, clearly no. That is the summary of our runs, but a summary hides the interesting part, so here is the call for each brief, the measure behind it and the builds to check it against. Both agents ran their vendor's strongest models on identical prompts, one attempt each.
Brief by brief: which agent did better?
- Landing page: Claude Code, on design. Its two pages look like two different brands; the two Codex pages are near-twins with the same fonts. Codex finished in 9 to 10 minutes against 28 to 43. Compare them.
- Analytics dashboard: Codex, on value. All four computed the same figures from the data and all controls worked. Codex delivered that for $0.54 to $2.79 at list prices against $12.80 to $13.63. Compare them.
- Physics simulation: a draw. All four ran without errors and kept every ball inside the hexagon. The Claude versions look richer; the Codex versions arrived in 8 minutes. Compare them.
- 3D shooter: Claude Code on atmosphere, Codex on finishing. Opus 5.5 hit our 150-minute limit and its recording ends in a game over; Fable 5.1 finished in 75 minutes with the most complete feel. Codex finished in 13 to 20 minutes, though GPT-6.1 Sol's player also died before the script ended. Compare them.
- Phone finance app: Claude Code, on feel. All four apps complete the send flow and update the balance correctly. Claude Code's add haptic feedback and animated confirmations; Codex's arrived in 9 to 14 minutes for $0.83 to $2.73 against 35 to 69 minutes and $11.46 to $21.06. Compare them.
- Open brief: Claude Code, on invention. Asked to invent a product, Claude Code wrote a night train and a sky lamp; both Codex builds invented a small gadget for using your phone less, in the same handwritten typeface. Compare them.
- Kanban board: Codex, on value. All four boards drag, undo, filter and open card details without a hitch. Codex built them in 12 to 15 minutes for $0.83 to $3.60; Claude Code took 45 to 101 minutes and about $13 to $20. Compare them.
What do the run logs say?
That the two agents work in opposite ways. On the first four briefs, Claude Code took 43 to 162 turns per build and ran 14 to 118 shell commands, many of them loading its own work in a headless browser. Codex ran each build as one turn of 10 to 26 commands.
| Model | Finished | Wall time | Output tokens | Files | Source size | Runtime errors | API-equivalent cost | Attempts |
|---|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 | Yes | 2h 30m | not recorded (stopped at the 150-minute limit) | 11 | 251 KB | 0 | — | 1 |
| Claude Fable 5.1 | Yes | 1h 14m | 239k | 9 | 165 KB | 0 | $20.48 | 2 |
| GPT-6 Astra | Yes | 19m 43s | 64k | 2 | 63 KB | 0 | $5.64 | 1 |
| GPT-6.1 Sol | Yes | 13m 19s | 61k | 3 | 65 KB | 0 | $1.03 | 1 |
Is Claude Code better than Codex on benchmarks?
Yes, by about six points, at a much higher cost. The public Terminal-Bench 4.0 leaderboard lists Claude Code with Opus 5.5 at 64.8% and Codex with GPT-6 Astra or GPT-6.1 Sol at 58.2%. The same table shows Opus 5.5 using about 8 billion tokens for its run against 1.5 billion for each GPT model, with run costs of $4.7k for Opus, $3.3k for Astra and $0.6k for Sol.
So which should you use?
Codex or Claude Code is a question of where your time goes. If you review and iterate, Codex's ten-minute first draft at under a dollar lets you take three swings before Claude Code finishes one. If you want to hand off a job and come back to something finished, Claude Code's hour of self-testing is worth it. I keep both: Codex for scaffolding and quick tools, Claude Code for anything a client will see. For how the two agents differ under the hood, read Claude vs Claude Code.
Settle it without logos: four shooters, two agents, names hidden.
Take the blind test
Frequently asked questions
Is Codex better than Claude Code?
Codex is better on speed and cost; Claude Code is better on depth and polish. In our tests Codex finished every build in 7 to 25 minutes for $0.54 to $7.27 at list prices, while Claude Code took 25 to 150 minutes and about $7 to $25 and produced more elaborate results.
Which agent scores higher on benchmarks?
Claude Code with Claude Opus 5.5 leads the public Terminal-Bench 4.0 leaderboard at 64.8%. Codex with GPT-6 Astra or GPT-6.1 Sol scores 58.2%, at a fraction of the token use and cost.
Should I use Codex or Claude Code for a new project?
Use Claude Code when you want the most complete first draft and can wait an hour. Use Codex when you want a working base in ten minutes and plan to iterate. Many developers keep both.
Were both agents given the same instructions?
Yes. The identical prompt, word for word, at maximum reasoning effort, one attempt each, with a 150-minute cap and no edits afterwards. Personal configuration files and plugins were switched off in both agents.