Codex vs Claude Code: Which to Use by Job, From 9 Same-Prompt Builds and Both Price Lists
Which to use for which job, from Pat Simmons's nine same-prompt builds in Claude Code and Codex, plus what each costs and which one runs out of usage first.
On this page
- Codex vs Claude Code: which to use for which job
- Is Codex better than Claude Code?
- Codex vs Claude Code for coding
- Claude Code vs Codex pricing: what each costs
- Which one runs out of usage first
- Which is easier if you don't write code
- How Pat tested them
- Update November 11, 2026: two new models Pat hasn't tested here
- Questions people ask
Use Claude Code with Opus 5.5 when the result has to look finished: in Pat Simmons's latest head-to-head, published September 25, 2026, it beat Codex running GPT-6 Astra on all three same-prompt builds.
Use Codex when speed, cost or usage limits weigh more, since Astra finished those builds in 30, 23 and 49 minutes against Opus 5.5's 85, 40 and 142, for less money each time. Codex also comes in a limited form with ChatGPT's free plan, while Claude Code starts with Claude Pro at $20 a month.
Last checked November 11, 2026: plan prices and what each plan includes, on OpenAI's Codex pricing page and Anthropic's Claude pricing page. The newest models, GPT-6.1 Sol in Codex and Anthropic's Fable 5.2, haven't been through Pat's head-to-heads yet. The build results come from his videos of July 11, September 5 and September 25, 2026.
Codex vs Claude Code: which to use for which job
Pick Claude Code with Opus 5.5 for one-prompt builds where the look and the small details count, and Codex with GPT-6 Astra for long jobs, a lower cost per build and the everyday engineering work Pat himself uses it for. Codex comes with ChatGPT plans and Claude Code with Claude's paid plans.
- Apps, sites and games from one prompt: Claude Code with Opus 5.5
- Big refactors and multi-step site architecture: Codex with GPT-6 Astra, Pat's day-to-day choice
- Very long unattended builds: Codex, which ran one for 10 hours 51 minutes without hitting Pat's limit
- Trying one without paying: Codex, through ChatGPT's Free plan
- Price at $20, $100 and $200 a month: level, and only Codex has a $500 plan
The winner has changed from round to round. Across Pat's nine same-prompt builds since July, the ChatGPT model in Codex won five, the Claude model in Claude Code won three, and one was split.
The full write-up of all nine Claude vs ChatGPT builds has Pat's notes on each.
| Video date | Build |  Claude Code model |  Codex model | Pat's pick | Claude Code time | Codex time | Claude Code cost | Codex cost |
|---|---|---|---|---|---|---|---|---|
| Sep 25, 2026 | Moby Dick scenes drawn only in code, in a risograph print style | Opus 5.5 | GPT-6 Astra | Opus 5.5 | 85 min | 30 min | $37.22 | $8.16 |
| Sep 25, 2026 | A remake of the Opus 5.5 launch film and web page, with sound made in code | Opus 5.5 | GPT-6 Astra | Opus 5.5 | 40 min | 23 min | $11.66 | $7.96 |
| Sep 25, 2026 | A Tony Hawk's Pro Skater-style game in a warehouse | Opus 5.5 | GPT-6 Astra | Opus 5.5 | 142 min | 49 min | $58.05 | $44.08 |
| Sep 5, 2026 | A 3D iPod app loaded with AI-made early-2000s songs | Fable 5.1 | GPT-6 Astra | GPT-6 Astra | About 1 hr 12 min to 1 hr 20 min | 37 min | $41 (Claude Code's own readout) | About $13 |
| Sep 5, 2026 | A 3D reconstruction of D-Day you can scrub through | Fable 5.1 | GPT-6 Astra | GPT-6 Astra | About an hour or less | 10 hr 51 min | Not given | Not given |
| Sep 5, 2026 | A Call of Duty-style shooter on the Nuketown map | Fable 5.1 | GPT-6 Astra | GPT-6 Astra | About 47 min | About 47 min | Not given | Not given |
| Jul 11, 2026 | A pixel-for-pixel clone of the Shots.so mockup app | Fable 5 | GPT-5.6 Sol | GPT-5.6 Sol | 30 min | 2 hr 51 min | $12 | Almost $100 |
| Jul 11, 2026 | An invented viral internet stunt and its drop website | Fable 5 | GPT-5.6 Sol | GPT-5.6 Sol | Under 20 min | About 30 min | $20 | About $5 |
| Jul 11, 2026 | A New York City learning platform from a deliberately rough brief | Fable 5 | GPT-5.6 Sol | Sol's design first; Fable's 3D map after one round of feedback | About 20 min | 11 min (first pass) | $95 | $11 |
Times are active build time. Costs are API-equivalent estimates: what each build's tokens (the small chunks of text a model reads and writes) would have cost at the model's published API price, worked out by an agent from the session logs.
Pat ran every build on a subscription, so nobody paid these amounts, and he says to take them with a grain of salt.
Is Codex better than Claude Code?
Claude Code is ahead as of Pat's latest test: running Opus 5.5, it beat Codex running GPT-6 Astra on all three builds on September 25, 2026. Three weeks earlier the result went the other way, when Astra in Codex beat Fable 5.1 in Claude Code on every build.
In July, Pat gave Fable 5 in Claude Code and GPT-5.6 Sol in Codex the same three builds. Sol won the Shots.so clone even though it took 2 hours 51 minutes to Fable's 30 and cost almost $100 to Fable's $12.
At the end of that video, Pat said "I really think Sol took the cake here", calling it more thorough, better at following directions and better at design (his July video at 25:46).
On the New York City platform he leaned toward Fable's 3D map once both had one round of feedback.
On September 5, GPT-6 Astra in Codex won all three builds against Fable 5.1, and Pat concluded that Astra was the best model for coding and that OpenAI had passed Anthropic for the first time. That was before Opus 5.5 came out.
On September 25, Opus 5.5 took all three, and Pat said "Opus 5.5 across the board seems to be the clear winner." He also preferred Fable 5.1's versions to Astra's in that round.
The skate game shows the gap: Opus 5.5 spent 142 minutes on it against Astra's 49, and its start screen lists keyboard and gamepad controls and a set of goals for the two-minute run.
Both skate games are live: click either start screen to play it. Pat's video of that round shows all three builds side by side.
Codex vs Claude Code for coding
For everyday engineering work Pat still reaches for Codex with GPT-6 Astra, and for one-prompt builds Claude Code with Opus 5.5 won his latest round.
At the end of the September 25 video (34:44), he said "I'm still partial to Astra in my day-to-day", naming "big refactors, large like multi-step site architecture" and 3D work, because "I really like how methodical Astra is."
He added that Astra "takes much longer, but it's much more thorough", that he rarely has to send it back to fix something, and that his builds are "not totally indicative of the more realistic everyday tasks" people do with these models.
Codex runs in the ChatGPT desktop app, a command-line tool, and an extension for VS Code, Cursor and Windsurf, and it also works inside Xcode and JetBrains editors. Codex Cloud gives each task its own workspace in OpenAI's cloud, where it keeps working while your computer is asleep, and Plus and above add automatic code review on GitHub.
Anthropic's Claude Code page lists a desktop app for macOS, Linux and Windows, plus the terminal, the web, iOS, Android, GitHub, VS Code, JetBrains and Slack.
Claude Code vs Codex pricing: what each costs
Both have paid plans at $20, $100 and $200 a month, and only Codex has a free tier: ChatGPT's Free plan and its $8 Go plan include limited Codex in the desktop app, while Claude Code isn't on Claude's Free plan. Codex also has a $500 plan with no Claude equivalent.
| Monthly price | Codex (ChatGPT plans) | Claude Code (Claude plans) |
|---|---|---|
| $0 | Free: Codex for quick tasks in the desktop app, on GPT-6 Luna, subject to rollout | Free: no Claude Code |
| $8 | Go: the same desktop-app Codex for lightweight tasks | No plan at this price |
| $20 | Plus: Codex on the web, in the command line, in the editor extension and on iOS, with automatic code review and Slack; GPT-6.1 Sol and GPT-6 Luna | Pro: $20 monthly or $17 a month billed yearly ($200 up front); Claude Code with Opus, Sonnet and Haiku; Fable paid with usage credits |
| $100 | Pro 100 | Max: 5 times Pro's usage, Fable up to half your weekly limit, $100 a month in API credits |
| $200 | Pro 200 | Ultra: 10 times Pro's weekly usage, no 5-hour limit for verified accounts, $200 a month in API credits |
| $500 | Pro 500, which adds Ultrafast mode | No plan at this price |
| Teams | Business: $20 per user a month billed yearly, $25 monthly, 2 users or more | Team: $20 per standard seat a month billed yearly, $25 monthly; premium seats $100, or $125 monthly |
| Pay as you go | Codex with an API key, at API prices, without the cloud features | Claude Code through the API, charged by tokens used |
Prices are in US dollars, from OpenAI's Codex pricing page and Anthropic's Claude pricing page, checked November 11, 2026. Anthropic's prices don't include tax. For more detail on each side, see every Codex plan and its limits and Claude Code pricing in full.
In Pat's Codex course from May, he said Plus "is plenty" if you're testing Codex and recommended the $100 Pro plan if it becomes your daily driver (his Codex course at 1:29).
For teams paying per use, Anthropic's Claude Code cost docs put the average across enterprise deployments at about $13 per developer per active day, or $150 to $250 per developer a month.
Per build, the cheaper tool depended on the model. In September, GPT-6 Astra in Codex cost less than Opus 5.5 in Claude Code on all three builds ($8.16 against $37.22 for the Moby Dick scenes).
In July it went the other way on the Shots.so clone, where GPT-5.6 Sol in Codex came to almost $100 against Fable 5's $12.
Pat said in July that he could check a session's cost in Claude Code with /cost or /usage, though the figures didn't always look accurate, and that he couldn't find a way to see it in Codex; OpenAI's pricing page now says /status in the Codex command line shows your remaining limits.
Which one runs out of usage first
In Pat's builds, Claude Code did. He hit his Claude usage limit partway through the Moby Dick build on September 25 and waited about 30 minutes for it to reset, while the same build used about 14% of his ChatGPT plan's allowance in Codex.
On September 5, GPT-6 Astra worked on the D-Day reconstruction in Codex for 10 hours 51 minutes and never hit a limit on Pat's $200-a-month ChatGPT Pro plan.
OpenAI's pricing page measures Codex on Plus and Business in five-hour windows, and weekly limits may also apply. Its estimates for Plus run from 5 to 45 local messages per five hours on GPT-6 Astra, 15 to 160 on GPT-6.1 Sol, and 350 to 3,000 on GPT-6 Luna. The Pro plans currently have no five-hour limit.
Fast mode uses your allowance 2.5 times as fast, Ultrafast 8 times, and ChatGPT Work shares the same allowance as Codex.
Plus and Pro users who run out can buy credits to keep going.
Claude's plans have a limit every five hours and every week. Max gives 5 times Pro's usage, and Ultra drops the five-hour limit for verified accounts and gives 10 times Pro's weekly usage.
Fable can use up to half the weekly limit on Max and Ultra, and on Pro it's paid with usage credits. Usage credits also let you keep working in Claude Code after you reach your plan's limit.
Which is easier if you don't write code
Pat hasn't compared the two for people who don't code, so neither wins here from use. Both have a desktop app where you describe what you want, and only Codex lets you try that for free.
Codex lives inside the ChatGPT desktop app for macOS, Windows and Linux, where you pick ChatGPT or Codex and describe the result you want.
Pat's Codex course is aimed at office work, and in it he says Codex "is fully agentic, meaning Codex can open your apps, it can edit your files, it can use your computer."
He sets it to full access so it can run commands without asking each time; the app also has more restricted permission settings.
Claude Code has a desktop app for macOS, Linux and Windows and needs Claude Pro or above. For a first build, see our walkthrough of building an app with Claude Code without writing code. If you're also weighing an AI code editor, see Cursor vs Claude Code.
How Pat tested them
Each round gave a Claude model and a ChatGPT model the exact same prompt, once, with no revisions, except the July New York City build, where Pat gave both one round of feedback.
Every build was deployed live to the internet. All models ran on high effort, the Claude models in Claude Code and the ChatGPT models in Codex, and Pat judged the results on camera by playing with each build. These were not blind rankings.
- September 25, 2026: Opus 5.5 and Fable 5.1 in Claude Code, GPT-6 Astra in Codex
- September 5, 2026: Fable 5.1 in Claude Code, GPT-6 Astra in Codex
- July 11, 2026: Fable 5 in Claude Code, GPT-5.6 Sol in Codex
The models changed every round, so the results compare a model running in a tool, and no round held the model constant to test Codex against Claude Code on their own. In July Pat was on a $100-a-month ChatGPT plan and a $200-a-month Claude plan.
Subscription plans don't show what a session would cost, so he had an agent estimate each build's API-equivalent cost from the session logs. The tests cover big creative builds only.
Update November 11, 2026: two new models Pat hasn't tested here
Anthropic has released Fable 5.2, which Pat reviewed on its own, and OpenAI's pricing page lists GPT-6.1 Sol for Codex on Plus and above. Neither has been in a Codex vs Claude Code head-to-head, so this page has no verdict on them yet.
Questions people ask
Is Claude Code free?
No. Claude Code isn't on Claude's Free plan; it starts with Pro at $20 a month, or $17 a month billed yearly.
Can Codex keep working while my computer is asleep?
Yes, in Codex Cloud: each task runs in its own workspace in OpenAI's cloud and keeps working while your computer is asleep.
Is Claude Max or ChatGPT Pro better value at $100 a month?
Pat hasn't compared them. Claude Max gives 5 times Claude Pro's usage and $100 a month in API credits, and OpenAI's page says its Pro plans currently have no five-hour limit in Codex.
Sources
- Codex pricing (OpenAI)
- Claude plans and pricing (Anthropic)
- Claude Code (Anthropic)
- Manage costs effectively, Claude Code docs (Anthropic)
- Codex IDE extension (OpenAI)
- Codex Cloud (OpenAI)
- ChatGPT desktop app (OpenAI)
- I Made GPT-6 Astra & Fable 5.1 Build the Same App (RAW RESULTS), Pat Simmons
- I Made GPT 5.6 and Fable 5 Build the Same App (RAW RESULTS), Pat Simmons
- Master 95% of Codex for Office Work (FULL COURSE), Pat Simmons
- Fable 5.2: NO-HYPE Full Review (& Testing), Pat Simmons
Comments
Sign in to join the conversation.