ZCode vs Claude Code: the free AI coder, benchmarked

A free app now scores within a few points of the tool developers pay for. Z.ai, a Chinese AI lab, released ZCode in early July. It builds working software from plain-English instructions, and on the one hard coding test every major lab reports, its model trails Claude's by about seven points while costing roughly a fifth as much. If you pay a developer, an agency, or a monthly subscription to build landing pages, automations, or internal tools, that gap matters.

Try ZCode if you want a free tool for small, self-contained builds and you care about cost. Stay on Claude Code if you run long, multi-step jobs or need the software to plug into everything else you use.

What these tools actually do

Both are agentic coding tools. You describe what you want in normal words ("build a booking page that texts me when someone fills it out"), and the tool writes the code, runs it, fixes its own errors, and ships it. Claude Code, from Anthropic, has been the one serious builders reach for. ZCode, from Z.ai, is the new free challenger: a desktop app running Z.ai's GLM-5.2 model. Same job. Very different price.

The scoreboard, in plain English

The fairest test here is SWE-bench Pro. It checks whether a model can fix real, messy bugs inside real software projects, not textbook puzzles. It is the one score all three of these models publish.

A quick note on trust. Claude Opus 4.8's number comes from Anthropic's published system card. GLM-5.2's comes from Z.ai's own docs. The Fable 5 figures are Anthropic's own claims and have not been independently checked yet, so read the top row as a promise, not a proof.

One independent test cuts the other way. On Code Arena, where people blind-vote on whose front-end code they prefer without knowing which model made it, GLM-5.2 ranked second. Ahead of Claude Opus 4.8. Behind only Claude's new top model, Fable 5. For everyday building, the cheap option is closer than the headline scores suggest.

Where the cheap one wins

Price, plainly. GLM-5.2 runs at about $4.40 per million tokens of output (roughly 750,000 words) against Opus 4.8's $25. The ZCode app costs nothing to download, and the Z.ai plan behind it starts near $18 a month against Claude Code's $20 entry. For a single page, a script, or a quick fix, GLM-5.2 gets close enough that the bill decides it.

Where Claude still wins

Long work. On tests of sustained, multi-step jobs, rewriting a whole project or running for hours, Claude Opus 4.8 pulls roughly twice as far ahead. The tool is deeper too. Claude Code runs in your terminal, code editor, browser, and phone, can run jobs in the background, and connects to your other software. ZCode is one desktop window, its background mode is not built yet, and its Linux version is still in beta. It is also a Chinese app, which raises data questions if you use the cloud version on client projects.

Worth watching

Z.ai ships updates weekly, and its cheaper model already runs inside Claude Code, so heavy users can keep the deeper tool and feed it the cheap model for routine work. Anthropic's Fable 5, out June 9, now sits above everything on the published scores. The distance between free and frontier is a few months now, not a few years. Our other tool breakdowns and the AI writing tools marketers actually use track where this goes next.