Trying it before you spend anything
OpenCode Zen
You want to see whether a terminal agent fits how you work, and you would rather not attach a card to find out.
At model rates
The vendor keeps ten models on its gateway at no charge for a limited time, and they are enough to answer the only question that matters at this stage. Read the terms on each before you paste anything real, because they differ: two run on providers that retain nothing and train on nothing, most of the rest are free while the lab behind them collects feedback, and NVIDIA's three carry an explicit instruction not to send personal or confidential data. For a scratch repository any of them is a fair deal; for client code, pick from the first two or pay.
The lowest bill that still has a ceiling you can see
OpenCode Go
You use an agent daily, you want one predictable charge, and you do not need a frontier model for most of it.
$10/mo
This is the most interesting thing the vendor sells, and not because of the price. The allowance is denominated in dollars of usage at a published rate card, set per model per month, with a fifth of it available in any five hours and half in any week — rather than in messages, credits or an unexplained window. That means you can work out in advance what you are getting, which is not true of any other subscription in this section. Two catches: the frontier models on the list carry the smallest allowances, and the plan is for single-member workspaces only. Go Plus, at four times the price, raises every allowance, by between two and eight times depending on the model.
A full day on real code with a frontier model
Your own API key
You are doing work that justifies the best available model and you want no ceiling in the way.
At model rates
Bring the key and pay the lab directly. opencode meters nothing of its own, so nothing is marked up and nothing is capped except by whoever issued the key — and the rate is the one on our own rate tables. This is also the route with no floor and no ceiling: an idle day costs nothing, and a heavy one costs whatever a heavy day of tokens costs.
You are already paying somebody for a model
ChatGPT Plus
There is a ChatGPT or Copilot subscription on your card already, and you would rather use it than open a second account.
$20/mo
The docs name a short list of subscriptions that work with zero setup, and this is the one most readers already hold. Turns spend the plan's allowance rather than an API balance, so the marginal cost of using opencode is nothing until you hit the limit you were already living with. Note the conspicuous absence from that list: a Claude subscription is not among them, and the reason is on the hub.
The code cannot leave your machine
Local model
Client data, an air-gapped box, or a policy that ends the conversation before a token reaches anyone's API.
Your hardware
An MIT client plus a local runtime is the only combination here that reaches a true zero, and the docs name five runtimes it supports. The bill moves rather than disappears: you pay in hardware, in electricity, and in the gap between a model that fits in your memory and one that does not. Worth it when the alternative is not sending the code at all.
A company that needs an answer for procurement
Enterprise
Several developers, an approved model gateway, and somebody who will ask where the code goes.
Contact sales
No published number, so treat the row as a starting point rather than a quote. What makes it worth a call is a term we have not seen elsewhere in this section: bring your own gateway and the vendor charges for seats but not for tokens. Combined with a central config that can disable every other provider, this is the rare enterprise story where the technical claim and the commercial one point the same way.