Is Claude good for coding?
Yes, and the evidence is in what it ships. On our briefs, Claude's builds are the most elaborate: a working booking calendar inside a landing page's hero, a dashboard chart annotated with its peak and its average, a neon-lit city around a shooter's arena. On the public Terminal-Bench 4.0 leaderboard, Claude Opus 5.5 in Claude Code leads at 64.8%, ahead of the GPT-6 models at 58.2%.
What makes the difference is behaviour, not just the model. Claude Code runs the model in a loop of writing, running and looking, and the Claude models use that loop heavily.
How do you code with Claude?
Pick the surface by how much you want Claude to touch.
| Way | What it does | What you need |
|---|---|---|
| Claude app | Writes code and previews pages as Artifacts | Free plan and up |
| Claude Code | Works in your project: edits files, runs commands, tests in a browser | Pro, from $17 a month billed annually or $20 monthly |
| Heavier agent use | Same, with five or twenty times Pro's usage | Max, from $100 a month |
| API | Pay per token inside your own tools | $4 and $20 per million tokens for Opus 5.5 |
Plan details are from Claude's pricing page and token prices from Anthropic's API pricing, both checked on October 21, 2026. Claude Code runs in a terminal, in VS Code and JetBrains, in a desktop app and on the web.
What did Claude build for us?
Every Claude Opus 5.5 build, with the facts from its run log. Fable 5.1's builds are on its model page.
| Brief | Finished | Wall time | Output tokens | Files | Source size | Runtime errors | API-equivalent cost | Attempts |
|---|---|---|---|---|---|---|---|---|
| SaaS landing page | Yes | 43m 20s | 144k | 1 | 132 KB | 0 | $7.15 | 1 |
| Analytics dashboard | Yes | 1h 21m | 285k | 1 | 115 KB | 0 | $13.63 | 2 |
| Balls in a spinning hexagon | Yes | 40m 11s | 211k | 1 | 31 KB | 0 | $8.73 | 1 |
| First-person shooter | Yes | 2h 30m | not recorded (stopped at the 150-minute limit) | 11 | 251 KB | 0 | — | 1 |
| Mobile finance app | Yes | 1h 8m | 295k | 1 | 115 KB | 0 | $21.06 | 4 |
| One-lane MOBA | Pending | — | — | — | — | — | — | 1 |
| Open brief | Yes | 1h 14m | 466k | 1 | 178 KB | 0 | $16.65 | 1 |
| Kanban board | Yes | 1h 40m | 446k | 1 | 113 KB | 0 | $19.98 | 2 |
Opus 5.5's 3D shooter is the one build in our first round that hit the 150-minute limit: Opus was still refining it, so the files on disk at that moment were published, and its recording ends in a game over at 25 seconds. Every other Claude build finished well inside the limit.
What is Claude especially good at?
Finishing the details. Three habits show up again and again in the run logs and the builds.
- It tests what it builds. Opus 5.5 ran 28 shell commands and 59 file reads for one landing page, about a dozen of them loading the page in a headless browser.
- It writes more. Its builds ran 2.5 to 9 times as many output tokens as the GPT builds of the same briefs, and its shooters carry up to four times as much source code. Both vendors' builds include accessibility basics such as ARIA labels and reduced-motion support; Claude's simply do more.
- It has a point of view. Opus 5.5 and Fable 5.1 produced visibly different landing pages from the same brief, where the two GPT models produced near-twins.
What kinds of coding is Claude best for?
Work where the first attempt matters. Mapped to our briefs:
- Client-facing pages. Claude's landing pages read as brands, not templates, and both Claude models produced a distinctly different design from the same brief.
- Interactive tools. Its dashboards add touches such as annotated peaks, average lines and a chart that flips into a table, all computed from the same data as everyone else's.
- Games and simulations. The most atmospheric results, with richer lighting and more interface, at the risk of running long: Opus 5.5 needed more than our 150 minutes for the shooter.
- Open briefs. When nothing is specified, Claude's habit of making choices becomes its biggest advantage.
Where the job is small and well defined, such as a script, a fix or a routine component, that same thoroughness mostly costs you time.
What does Claude's self-testing actually look like?
A loop of building, serving, looking and fixing. In the run log for its 3D shooter, Claude Opus 5.5 issued 118 shell commands: 37 involved a local web server so it could load its own game, and 42 involved Playwright, a tool for driving a headless browser, to play it and capture what it rendered. It read files 55 times and rewrote them 13 times. Fable 5.1 did the same kind of loop with fewer steps, 42 shell commands for its shooter.
You never see that loop in the chat app, because there Claude cannot start a server or open a browser on your machine. It is the main reason Claude Code's output differs from what you get by pasting a prompt into a chat, as the Claude vs Claude Code comparison explains.
How much did coding with Claude improve with the 5.5 generation?
Noticeably, on the public numbers. The Terminal-Bench 4.0 leaderboard lists Claude Opus 5 at 53.9% and Claude Fable 5 at 44.5%, against 64.8% for Opus 5.5 and 57.9% for Fable 5.1. Anthropic also cut prices: Opus 5.5 costs $4 and $20 per million input and output tokens, 20% less than Opus 5, and its cached reads are 60% cheaper, according to the Opus 5.5 announcement. For agent work, which is mostly cached reads, that is the number to watch.
Where does Claude fall short?
Knowing when to stop, and what that costs. Claude's builds took 3 to 11 times as long as the GPT builds and wrote 2.5 to 9 times as many output tokens. Opus 5.5 read about 10 million input tokens to build one landing page, most of them cheap cached reads but still billed. That appetite also hits subscription limits: two maximum-effort Claude Code sessions side by side exhausted a five-hour usage window in about two hours during our runs.
More code is not always better code, either. Opus's shooter is 11 files and 251 KB; the GPT shooters are 2 to 3 files and about 64 KB. If you will maintain the result yourself, Claude's version is more to learn.
Opus 5.5 or Fable 5.1 for coding?
Opus 5.5 by default. It costs $4 and $20 per million input and output tokens against Fable's $10 and $50, Anthropic's own model guide recommends starting with it, and it leads the public leaderboard. In practice the two cost about the same per build on our briefs, because Opus works longer: $7.15 to $13.63 for Opus and $8.75 to $12.80 for Fable on the three briefs both finished inside the limit. Choose Fable when you want a calmer, more restrained result, or for long-horizon work where Anthropic positions it.
How do you get better code out of Claude?
Give it scope and a budget. Claude will fill whatever room you leave, so say what is out of scope and how long it should spend. Run one heavy session at a time if you are on a subscription. Keep maximum effort for the hard builds and use a lower effort for routine edits; Anthropic notes that Fable 5.1 defaults to high effort in Claude Code. And use the same brief shape as ours: fixed content, a feature list, technical constraints and a phone requirement. Any of our brief prompts is a good template.
Frequently asked questions
Is Claude good for coding?
Very. In Claude Code, Claude Opus 5.5 and Claude Fable 5.1 built every project we gave them and shipped the most elaborate results of any models we tested. The cost is time and tokens: their builds took 25 to 150 minutes and about $7 to $25 each at list API prices.
Which Claude model is best for coding?
Claude Opus 5.5 for most work. Anthropic recommends it as the default, it leads the public Terminal-Bench 4.0 leaderboard at 64.8%, and it is the cheaper of the two per token. Use Claude Fable 5.1 for long, demanding tasks where Opus falls short.
Can I use Claude for coding for free?
You can ask the free Claude app for code and preview pages as Artifacts. Claude Code, the agent that edits files and runs commands in your project, needs a paid plan, from $17 a month billed annually for Pro.
Why does Claude take so long to code?
Because it checks its own work. On our briefs Claude Code ran 9 to 118 shell commands per build, many of them loading the result in a browser to see what it rendered and fix it. That loop costs time and tokens and is the reason its builds are the most finished.