I've been using OpenCode on my secondary computer for a while now. I also use Codex, but on that machine OpenCode makes more sense to me because I have much more freedom to choose what model I want to use without changing the whole tool.
Most of the time I use Big Pickle. I've found it surprisingly smart for a free model. I normally use OpenCode for research, automation and coding tasks — nothing crazy — and Big Pickle has been good enough for most of that. It knows how to use tools, can work through normal programming tasks and usually understands what I'm trying to achieve without me having to explain everything several times.
This is probably what I like most about OpenCode. I don't need a powerful computer to use a powerful model.
My secondary computer doesn't have to run the model itself. OpenCode runs there, while the actual inference can happen through Zen, Go or another provider. If I want to try another model, I can just change it. I can also use OpenAI models through OpenCode instead of having to move to a completely different coding agent.
I still like local AI and there are obvious reasons to run models yourself, but sometimes I simply want to get something done. Running a local model also means having enough RAM or VRAM, choosing the right quantization, dealing with context size and accepting that a model small enough to run on an older computer may not be the model I actually want to use.
OpenCode avoids that problem for me.
It has been more stable than I expected
OpenCode feels very polished for an open-source project and it gets updates very frequently. Stability has actually surprised me because I expected more problems considering how quickly the project is moving.
That doesn't mean I haven't found any.
One issue I've noticed a few times happens when I give it several tasks to complete. It starts working through them normally, but later I notice that something from the original task list was never finished. Sometimes it looks like the agent simply decided the work was done before everything actually was.
I still don't know if this is an OpenCode issue or a model issue. Since I normally use Big Pickle, blaming OpenCode without testing the same task with another model wouldn't tell me much. It could also be related to context or simply the model losing track of the original instructions during a longer task.
I need to test that properly.
For the type of work I normally do with it, though, OpenCode has been stable enough that I keep using it. Research works well, automation is useful and for normal programming tasks I haven't felt that I'm constantly fighting the application.
I also like that I'm not stuck with whatever model the developer of the agent decided I should use. If Big Pickle isn't doing well with something, I can switch. If I want to use one of the models available through Go or Zen, I can do that too.
Codex is different. I use it and I like it, but it's a much more controlled experience around OpenAI's own models. On my secondary computer I prefer having OpenCode available because I can experiment more.
I don't see that as OpenCode replacing Codex. At least not yet. They just fit differently into the way I work.
Free models are useful, but I would be careful with the data
One thing I wouldn't ignore is why some of these models are free.
OpenCode currently says that while Big Pickle is being offered for free, data collected from its use may be used to improve the model. There are other free Zen models with similar exceptions. Most of the normal paid Zen models are covered by different retention and training conditions, so this isn't something I would generalize to every model available through OpenCode.
For me this mainly changes what I'm willing to send.
If I'm testing something, researching a topic or working with code that isn't sensitive, I'm comfortable using Big Pickle. I wouldn't send it passwords, API keys, personal information, private company information or anything else I wouldn't want included in data used to improve a model.
You have to check the conditions of the model you're actually using. Seeing the word free is not enough.
This is also why I don't think cloud models and local models are interchangeable. A local model gives me much more control over where the data goes, but it also requires hardware capable of running it. OpenCode gives me another option: use the hardware I already have and access much stronger models remotely.
For someone who wants to experiment with AI agents but doesn't have a computer capable of running a good local model, I think OpenCode is one of the easiest ways to start.
Install it on the computer you already have, try different models and see what is actually useful to you before spending money on hardware.
That's more or less what OpenCode has become for me. On my secondary computer I can research, automate things and do normal coding work without caring whether that machine could ever run the model locally.
And right now Big Pickle is good enough that I use it much more often than I expected.