← All posts

Claude Code vs OpenCode, Codex and Chinese AI Models

AI coding tools have gone from "neat autocomplete" to "go fix this bug while I make coffee" surprisingly fast. The problem now isn't whether to use one, it's which one. There are a lot of options, they all look great in demos, and they all cost money in slightly different ways.

I've been poking at most of them, so here's my honest take on where things stand - at least in October 2026.

The Main Contenders

Claude Code is Anthropic's terminal-based coding agent. You point it at your project, tell it what you want in plain English, and it reads your files, edits code, runs commands and checks its own work. It asks before doing anything risky, which I like. It feels like pair programming with someone who actually read the whole codebase first.

OpenAI's tools (Codex and the GPT models behind ChatGPT) are the obvious rival. Codex is more of a "hand it a task and come back later" setup, often running in the cloud on its own copy of your repo. It's fast, it's generally cheaper per task, and if you already pay for ChatGPT it's sitting right there in your subscription. For well-defined chunks of work it's really good.

OpenCode is the open-source option, and it's the one I find most interesting philosophically. It's a terminal agent like Claude Code, but it isn't tied to one company. You bring your own model, whether that's Claude, GPT, something running locally or one of the Chinese models. No lock-in, and you pay for whatever you actually use.

The Chinese Models Are Getting Scary Good

This is the part I didn't expect a year ago. The newer models out of China (think Kimi, GLM, Qwen, DeepSeek and MiniMax) have gotten genuinely good at coding. Not "good for the price", just good.

A few things make them appealing:

  • They're dramatically cheaper, often a fraction of what the big Western labs charge per token.
  • Many have open weights, so you can self-host them or pick whichever provider is cheapest.
  • They plug straight into tools like OpenCode, so trying one costs you a config change.

Kimi (from Moonshot) is the one I hear about most for agentic work. It's a huge open-weight model built to plan, call tools and keep going over long tasks, and the company now ships a version tuned specifically for coding. Its own numbers still put it a little behind the top Western models, but close enough that it's a real option and not just a curiosity.

DeepSeek is the price story. Its API is OpenAI-compatible, so it plugs into almost anything, and the rates are tiny: a "Flash" tier that costs pennies per million tokens and a bigger "Pro" tier for harder coding and agent work that's still only a couple of dollars. Cached input is cheaper still, which matters a lot when an agent keeps re-reading the same project files. Pricing has been shifting (it now varies by time of day), so check before you budget around it.

GLM, Qwen and MiniMax are worth a look too, mostly because of how they're sold. GLM (from Z.ai), MiniMax and Alibaba's Qwen all offer flat monthly "coding plans" instead of pay-per-token, and they publish guides for using them inside Claude Code and OpenCode. Usually that's just a couple of environment variables or a built-in connect command. A flat fee is a nice way to stop worrying about the meter running, though read the fine print, since some plans only allow use through approved tools.

There are real trade-offs. Almost every benchmark you'll see is published by the company that made the model, and independent reviews are still thin. Some of these models also burn extra "thinking" time on simple tasks, and depending on the provider, your code may be going to servers you'd rather it didn't, so think about what you're sending. Self-hosting the open weights solves that, if you have the hardware.

Are they at the very top yet? On the hardest, messiest tasks, in my experience, not quite. But the gap is shrinking every few months, and benchmark scores will tell you they're closer than ever. I'd take any leaderboard with a grain of salt though. Your own codebase is the only benchmark that matters.

Why I Still Pick Claude Code

Here's my personal verdict: Claude Code is the one I'd choose today.

The biggest reason is that the code it writes is just less error-prone. I find myself fixing fewer mistakes, and I spend way less time going back and forth saying "no, not like that" or "you broke the thing next to it." It tends to get the change right, or close to right, on the first or second try. When you're paying with your own time, that matters more than any benchmark.

It's also good at working inside an existing project. It follows the patterns that are already there instead of inventing its own style, which is exactly what you want on a real codebase and not a fresh toy project.

It's Surprisingly Good at Design Too

Here's something that doesn't get talked about enough: Claude is a pretty capable designer. I don't mean it replaces a real designer with a good eye, but for a developer who just wants a page that doesn't look like 2009, it's a big deal.

Describe a layout in plain English and you get clean, sensible markup and styling, usually with decent spacing, a consistent type scale and sane colors. It's comfortable with Tailwind and component frameworks like Svelte, so the result lands in your codebase and not in some throwaway mockup. You can also hand it a screenshot or a rough sketch and ask it to build something similar, then keep nudging it with things like "tighten the spacing" or "make the cards feel less cramped." It also tends to remember the boring-but-important stuff like responsive layouts, dark mode and accessibility, which are exactly the things I'd forget on a first pass.

It won't have taste out of the box, so you still need to steer. But getting from a blank page to a solid first draft in minutes, then refining by conversation, is a huge speedup. That's a big part of why I find it more useful than tools that only shine at the pure logic side.

The Catch: It's Expensive

I won't pretend otherwise. Claude Code is not cheap. Between the subscription tiers and what you burn through on heavy days, it's easily the priciest option on this list.

For me it's been worth it, because the time it saves outweighs the bill. But "worth it" is a moving target, and I'm honestly not sure I'll renew. The Chinese models look really promising, and if I can get 90% of the quality at 10% of the cost through something like OpenCode, that's a hard deal to ignore. I'll probably keep an eye on them and run a few side-by-side tests on my own projects before deciding.

So if you're picking one today and your budget allows it, Claude Code is my recommendation. If money is tight or you hate lock-in, start with OpenCode and a cheaper model. And whichever way you go, check back in a few months, because this space moves fast enough that this whole post might be out of date by then.