What is the difference between Claude and Claude Code?
Claude is the model, Claude Code is the workshop. When you use Claude on the web, desktop or phone, you send a message and get an answer, which can include code or a previewable Artifact. Claude Code, in Anthropic's words, "reads your codebase, edits files, runs commands, and integrates with your development tools". It runs in a terminal, in VS Code and JetBrains, in a desktop app and in the browser.
The practical difference is who does the loop of write, run, look and fix. In the chat app you do it: copy the code, run it, report what broke. In Claude Code the model does it, as many times as it decides it needs to.
Where does each run and what does each cost?
They are sold together but unlock at different tiers.
| Claude (the app) | Claude Code | |
|---|---|---|
| Where it runs | Web, desktop, iOS, Android | Terminal, IDE, desktop app, browser |
| What it can touch | The conversation and its Artifacts | Your files, your shell, your browser |
| Free plan | Yes | No |
| Included from | Free | Pro: $17 a month billed annually, $20 monthly |
| Heavier use | Max, from $100 a month | Max, from $100 a month |
Plan details are from Claude's pricing page, checked on October 21, 2026. Developers can also run Claude Code against the API and pay per token at the rates on Anthropic's pricing page: $4 and $20 per million input and output tokens for Opus 5.5, $10 and $50 for Fable 5.1.
Does Claude Code use a different model?
No, but it can use it differently. The models are the same: Anthropic's model overview lists one family across its products, although which models your plan includes varies. The defaults are not the same, though. Anthropic's Fable 5.1 announcement notes that Fable 5.1 defaults to high effort in Claude Code and to medium effort on Claude.ai. Effort is how long the model is allowed to think, and on hard work it shows. Every build on this site ran at maximum effort.
What does the harness change in practice?
Almost everything about how the work gets done, and our run logs make that concrete. Claude Code took between 43 and 162 turns per build on the first four briefs, where a turn is one round of acting and observing. A large share of those turns were the model opening its own page in a browser, looking at the result and fixing it.
- Claude Opus 5.5 used 91 turns on the landing page, 118 on the dashboard and 162 on the 3D shooter, where it was still working when our 150-minute limit arrived.
- Claude Fable 5.1 used 43 to 65 turns per brief and finished every one, the longest in 75 minutes.
- Input tokens tell the same story. Opus read about 10 million input tokens to build a 132 KB landing page, because every turn re-reads the growing project. Most of that is billed at the cheap cached-read rate.
For comparison, on those briefs OpenAI's Codex ran each GPT build as one long turn of 10 to 26 shell commands. That is a design difference between the two agents rather than proof that one tested more carefully, and the Codex vs Claude Code comparison on this shelf goes into it.
| Model | Finished | Wall time | Output tokens | Files | Source size | Runtime errors | API-equivalent cost | Attempts |
|---|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 | Yes | 1h 21m | 285k | 1 | 115 KB | 0 | $13.63 | 2 |
| Claude Fable 5.1 | Yes | 38m 43s | 142k | 1 | 77 KB | 0 | $12.80 | 1 |
| GPT-6 Astra | Yes | 10m 27s | 35k | 1 | 68 KB | 0 | $2.79 | 1 |
| GPT-6.1 Sol | Yes | 7m 26s | 36k | 1 | 67 KB | 0 | $0.54 | 1 |
What does Claude Code actually do between your prompt and the result?
It reads, runs and looks, many times over. We counted the tool calls in each run log. On the landing page, Claude Opus 5.5 read files 59 times and ran 28 shell commands, about a dozen of which started a headless browser with Playwright to load the page and capture what it rendered. On the 3D shooter it ran 118 shell commands, 37 of them involving a local web server so it could load and play its own game. Claude Fable 5.1 worked more economically, with 14 to 42 shell commands per brief on the first four briefs.
None of that is visible in the chat app, where the model cannot start a server or open a browser on your machine. This is the real answer to whether the harness changes what gets built: the model in Claude Code gets to see its own mistakes before you do. In the logs you can watch it load a page, read what rendered, and edit the file again, with no human involved.
What does it mean for your usage limits?
Agent work burns through subscription limits much faster than chat. Every turn re-reads the project, which is why Opus 5.5 read about 10 million input tokens for one landing page. We learned the practical limit the hard way: on October 17, 2026, two maximum-effort Claude Code sessions running side by side exhausted a subscription's five-hour usage window in about two hours and were stopped with a "session limit" error until it reset. Both interrupted attempts are listed on their builds. If you plan long agent sessions, budget for a Max plan or API billing, and run one heavy session at a time.
When should you use Claude, and when Claude Code?
Use the Claude app when the output is an answer: an explanation, a plan, a review of code you paste, a one-off script or a single page you want to preview. It is free to start and nothing touches your machine.
Use Claude Code when the output is a project. Anything with several files, anything that has to run, anything you want tested before you see it. That is where the extra steps pay off, and it is the only way the builds on this site could have been made without a human in the loop.
My own rule: if I would have to copy the answer into a file and run it myself, it was a Claude Code job.
Which Claude model should you run inside Claude Code?
Start with Claude Opus 5.5, which Anthropic's model overview recommends for most workloads, and move to Claude Fable 5.1 for long, demanding work where Opus falls short. Our runs add two facts to that advice. Opus worked longer than Fable on every brief, 40 to 81 minutes against 25 to 39 on the three briefs both finished inside the limit, because it tested more. And despite Opus's lower token price, the two cost about the same per build at list prices: $7.15 to $13.63 for Opus and $8.75 to $12.80 for Fable on those briefs. Pick on the results you prefer, not on the price sheet.
How can you see the difference yourself?
Pick one of our briefs, paste the exact prompt into the Claude app, and compare its Artifact with the Claude Code builds on that brief. The prompt is identical; only the harness differs. You will see the same model's taste in both, and you can judge for yourself how much of the brief works when you click through each.
Frequently asked questions
Is Claude Code a different model from Claude?
No. Claude Code runs the same Claude models, such as Claude Opus 5.5 and Claude Fable 5.1. What differs is the harness around the model, its tools and its defaults. Anthropic notes, for example, that Fable 5.1 defaults to high effort in Claude Code and medium effort on Claude.ai.
Is Claude Code free?
No. Claude Code is not part of the free plan. It comes with Claude Pro, from $17 a month billed annually or $20 billed monthly, with Max plans from $100 a month, according to Claude's pricing page. It can also be used with pay-as-you-go API billing.
Can the Claude chat app build a website or app too?
It can write one and preview it as an Artifact, which works on the free plan. What it cannot do is run commands on your computer, open the result in a browser and fix what it finds over many steps, which is what Claude Code did for every build on this site.
Should I use Claude or Claude Code for coding?
Use the Claude app to think, explain, review a snippet or sketch a single page. Use Claude Code when the work lives in files: a real project, a multi-file feature, anything that needs to be run and tested.