Can ChatGPT really code a whole project?
Yes, from a single prompt, when it runs as an agent. We did not paste code back and forth in a chat window. We gave each brief to Codex, OpenAI's coding agent, with GPT-6 Astra or GPT-6.1 Sol at extra-high reasoning, and let it work alone with a 150-minute limit. Every build finished, in one uninterrupted turn of 10 to 33 shell commands, far inside the limit.
The results are not demos. The landing pages have working pricing switches, FAQs and phone menus. The dashboard computes exactly the figures its dataset implies, $155,149.80 in revenue from 2,109 orders. The shooter has waves, reloading, a HUD and a game-over screen.
How do you code with ChatGPT?
Three ways, from lightest to heaviest.
| Way | What it does | What you need |
|---|---|---|
| ChatGPT chat | Writes code you copy into your own files | Any plan, including Free |
| ChatGPT Sites | Builds and hosts a site or small app from a conversation | A ChatGPT plan with Sites |
| Codex | Works in your project: writes files, runs commands, checks its work | A paid ChatGPT plan, or API billing |
ChatGPT Plus costs $20 a month, according to OpenAI's Help Center, and its allowance can be used in participating tools through Sign in with ChatGPT. The Pro tiers, at $100, $200 and $500 a month, list Codex among their features, per OpenAI's Pro plans page. Developers can also pay per token: $2 and $10 per million input and output tokens for GPT-6.1 Sol on OpenAI's model page.
What did ChatGPT's models build for us?
Here is every GPT-6.1 Sol build with the facts from its run log. GPT-6 Astra's builds are on its model page and look remarkably similar.
| Brief | Finished | Wall time | Output tokens | Files | Source size | Runtime errors | API-equivalent cost | Attempts |
|---|---|---|---|---|---|---|---|---|
| SaaS landing page | Yes | 9m 08s | 45k | 1 | 73 KB | 0 | $0.69 | 1 |
| Analytics dashboard | Yes | 7m 26s | 36k | 1 | 67 KB | 0 | $0.54 | 1 |
| Balls in a spinning hexagon | Yes | 7m 54s | 43k | 1 | 25 KB | 0 | $0.65 | 1 |
| First-person shooter | Yes | 13m 19s | 61k | 3 | 65 KB | 0 | $1.03 | 1 |
| Mobile finance app | Yes | 13m 46s | 48k | 1 | 57 KB | 0 | $0.83 | 1 |
| One-lane MOBA | Yes | 19m 41s | 67k | 3 | 85 KB | 0 | $1.12 | 1 |
| Open brief | Yes | 10m 51s | 40k | 1 | 65 KB | 0 | $0.71 | 1 |
| Kanban board | Yes | 14m 31s | 53k | 1 | 64 KB | 0 | $0.83 | 1 |
At list API prices, Sol's builds cost $0.54 to $1.12 each and Astra's $2.12 to $7.27. Both are a fraction of the roughly $7 to $25 the Claude builds of the same briefs would have cost.
What is ChatGPT good at in coding?
Finishing. Both GPT models reliably delivered everything the brief listed, in one pass, without wandering off. Three strengths stand out in the logs:
- Speed. Builds took 7 to 25 minutes, against 25 to 150 minutes for the Claude models on the same briefs.
- Economy. Fewer commands, fewer tokens and a smaller, more readable codebase: the GPT shooters are 2 to 3 files and about 64 KB, the Claude shooters 9 to 11 files and up to 251 KB.
- Correctness on data. Where a brief had numbers to compute, the GPT builds computed them right.
What kinds of coding is ChatGPT best for?
Work with a clear finish line. Mapped to our briefs, the pattern is consistent.
- Internal tools and dashboards. The best value of anything we tested: correct numbers, sortable and filterable tables, a working dark mode, in about ten minutes and well under a dollar with GPT-6.1 Sol.
- Marketing pages. Complete and responsive, but generic. Expect a template-quality page that you will want to give a personality.
- Simulations and algorithms. Solid. Both GPT physics simulations ran their 20 seconds without errors, and all twenty balls are still inside the hexagon in the final frame.
- Games. Playable, with every required system present, but flatter and blockier than the Claude games, and less robust under our scripted input.
Where the job is open-ended, such as "make it feel premium" or "invent something memorable", ChatGPT's tendency to pick the safe default matters more, and a single prompt gets you less.
Is ChatGPT's code easy to maintain?
Easier than most. The GPT builds are small and direct: the landing pages are single files of 64 KB and 73 KB, and the shooters are two or three files of about 64 KB in total, against nine to eleven files and up to 251 KB for the Claude shooters. Fewer moving parts make the code faster to read and safer to change by hand.
Smaller does not mean careless. On the landing page and dashboard briefs, all four builds carry ARIA labels, keyboard handling and reduced-motion support, and GPT-6.1 Sol's landing page has more ARIA attributes than any other. What the GPT builds leave out is scope, not care: fewer features, fewer effects, less to maintain.
How much did coding with ChatGPT improve with GPT-6?
Substantially, by both OpenAI's numbers and independent ones. The public Terminal-Bench 4.0 leaderboard lists GPT-5.6 Sol at 37.3% and the GPT-6 models at 58.2%. OpenAI also reports that it sped up the Codex harness alongside Astra, so that tasks on the Mind2Web benchmark finish 1.9 times faster than with GPT-5.6 Sol, and that Astra stayed within its authorised scope in every case of a test where GPT-5.6 Sol overstepped 48% of the time, according to its launch post. For an agent you leave alone with your code, that last number may matter most.
Where does ChatGPT fall short?
Variety and ambition. Ask GPT-6 Astra and GPT-6.1 Sol for the same thing and you often get the same answer: their landing pages share a headline structure, an accent colour and both fonts, their physics demos share a name, and their shooters both put a door marked Relay in front of the player. The designs are tidy and safe. If you want a distinctive look, you will need to ask for it explicitly, or ask for two contrasting directions.
The GPT agents also test their work less than the Claude agents do, which is part of why they are fast. In our shooter, Sol's player died 35 seconds into our 40-second input script while Astra's survived. On the public Terminal-Bench 4.0 leaderboard the GPT-6 models score 58.2% against 64.8% for Claude Opus 5.5.
Which ChatGPT model should you use for coding?
GPT-6.1 Sol for nearly everything. It ties Astra on Terminal-Bench, finished every one of our builds faster than Astra, and costs a fifth as much per token. Keep GPT-6 Astra for the cases where you have watched Sol fail, and say so in your prompt when you switch; on our briefs, the switch did not change much.
How do you get better code out of ChatGPT?
Write the brief like ours. Every brief on this site that produced a working GPT build had the same shape: the exact content and data, the sections or features as a list, technical constraints such as "a single HTML file that opens from disk", and a requirement to work on a phone. Then add the two things the GPT models will not do on their own: ask for a distinctive visual direction, and ask Codex to open the result in a browser and fix what it finds before it finishes. You can copy any of our prompts from the brief pages as a starting point.
Frequently asked questions
Is ChatGPT good for coding?
Yes. The GPT-6 models behind ChatGPT built every project we gave them, from a landing page to a top-down MOBA, in 7 to 25 minutes each with no runtime errors in our recordings. Their code is complete and clean; their designs are conservative and often alike.
Can I use ChatGPT for coding for free?
You can ask for code in the free ChatGPT app. The agent that writes and runs code in your project, Codex, comes with paid plans: OpenAI lists Codex among the features of its Pro tiers, which start at $100 a month, and Plus costs $20 a month with usage in participating tools through Sign in with ChatGPT.
Which ChatGPT model is best for coding?
GPT-6.1 Sol. It ties GPT-6 Astra on the public Terminal-Bench 4.0 leaderboard at 58.2%, finished each of our builds faster than Astra, and costs a fifth as much per token at API prices.
How good is ChatGPT at coding compared with Claude?
Faster and cheaper, less elaborate. On our briefs the Claude models worked 3 to 11 times longer and produced richer builds; on the public Terminal-Bench leaderboard Claude Opus 5.5 scores 64.8% against 58.2% for the GPT-6 models.