OpenCode is one of the few tools in this tracker where the headline price is genuinely zero and stays that way. It’s an MIT-licensed terminal coding agent, and nothing about installing it, running it, or using every feature it ships requires an account. What you pay for is inference, and OpenCode gives you three separate ways to pay for that — one of which routes no money and no data through OpenCode at all.
That first option is the default and the one the docs keep pointing back to. Bring an API key you already have, from Anthropic, OpenAI, Google, Bedrock, Azure, Groq, OpenRouter or anywhere else, and OpenCode sends requests straight to that provider at that provider’s rates. There’s no subscription sitting in between and no markup. Point it at a local model through Ollama and the marginal cost is your electricity. Both paid tiers below are OpenCode acting as a model provider, and the documentation describes each as “completely optional” in those words.
OpenCode Zen is the gateway option: a curated list of models the team has tested and benchmarked against specific providers, on the argument that the same model served by different providers gives you meaningfully different quality. You add billing details, get an API key, and are charged per request against prepaid credits, at per-1M-token rates published model by model. There’s no published minimum top-up, so we’re not quoting one. Zen also carries a rotating set of genuinely free models, which is where the pricing question turns into a privacy question — more on that below.
OpenCode Go is the $10/month subscription, and it’s the most interesting piece of pricing design here. Instead of counting requests or messages, Go denominates its limits in dollars of usage: $12 per rolling 5 hours, $30 per week, $60 per month. That means your effective request count depends entirely on which model you pick, and OpenCode publishes the estimates rather than making you guess — a Grok 4.6 session gets you roughly 169 requests per 5 hours, while MiMo-V2.5 gets you about 30,100. The stated goal is to return about 6x the subscription price in usage for most models, funded by bulk discounts and reserved GPU capacity. That multiplier isn’t uniform, and the docs are unusually candid about why: for models that are new or already cheaply priced, they couldn’t negotiate a discount, so the included monthly usage drops as low as $15 against $100 for others. One constraint that’s easy to miss: only one member per workspace can subscribe to Go, so it isn’t a team plan wearing a cheap price tag.
Run past those limits and requests stop, unless you’ve enabled the “Use balance” option in the console, at which point Go spills over onto your prepaid Zen credits. It’s opt-in, which is the right default for something that turns a fixed $10 into a variable bill.
The part worth reading carefully before you pick a tier is what each model does with your data, because OpenCode publishes it per model and the answer is not uniform. On Go, most models are documented as “not used” for training with 0 days of retention. Grok 4.6 and GPT 5.6 Luna sit at 30 days, for reasons the docs attribute to features that depend on stored data and to abuse-monitoring logs respectively. The two Muse Spark Contributor models are listed as trained on and explicitly not zero-retention — that’s the trade being made for the discount, and OpenCode says so plainly rather than burying it.
The same asymmetry runs through Zen’s free models. OpenCode’s privacy section states that its providers follow a zero-retention policy and don’t train on your data, then names the exceptions: Big Pickle, MiMo-V2.5 Free and Ling 3.0 Flash Fin Free may use collected data to improve the model during their free period, and NVIDIA’s free Nemotron endpoints are labelled trial use only, with a direct instruction not to submit confidential data and a note that sessions are logged. So “free” on Zen means free of charge, not free of conditions, and the conditions are concentrated in exactly the tier a cost-conscious developer would reach for first. If you’re working on anything under an NDA, the BYO-key path — where OpenCode never sees the request at all — is the only configuration that removes the question entirely.
Comparing this to seat-priced competitors is awkward on purpose. There is no seat here, no per-user minimum, and no enterprise tier gating the agent’s features behind a sales call. Team workspaces exist and are currently free as part of a beta, with pricing described as still to be announced — which is worth treating as a temporary $0 rather than a commitment. For an individual developer, the honest framing is that OpenCode’s own pricing is a floor of nothing plus whatever inference you choose, and the $10 Go tier is best read as a bet that you’d rather have a predictable ceiling on open models than an unpredictable bill on frontier ones.