OpenCode is an open-source coding agent, and the team behind it runs two optional services around it: Zen, a pay-as-you-go model gateway, and Go, a flat monthly subscription for open coding models. Both also host a rotating set of free models, which is why the OpenCode page is one of the busiest on LLM Perks. This guide explains how Zen and Go differ, what the free models really cost you in data and availability, how Go's usage limits work, and how to choose. Everything below comes from OpenCode's console documentation, read on November 13, 2026, and from our own catalog, which re-reads the public Zen model list on every check.
Zen and Go in one table
| Zen (pay as you go) | Go (subscription) | |
|---|---|---|
| How you pay | Add credits; you are charged per request | $10 per month (Go), $40 per month (Go Plus), $20 per member per month (Go Teams) |
| Models | Frontier models from OpenAI, Anthropic, Google and others, plus open models, all priced per 1M tokens | A curated list of open coding models, for example GLM, Kimi, Qwen, MiniMax, MiMo, DeepSeek, Grok and Step |
| Limits | Your credit balance and optional budgets | Dollar-denominated allowances per period |
| Best for | Variable use, access to frontier models | Daily agent use at a predictable price |
| Required? | No. OpenCode works with any provider | No. "Completely optional" |
Both work like any other provider inside OpenCode: sign in to the console, add billing details, copy your API key, run /connect in the terminal UI, and run /models to choose a model.
Zen: pay-as-you-go access
OpenCode describes Zen as a list of models tested and verified by the OpenCode team. You top up credits and pay per request, at per-1M-token prices published in the docs. Key features from the documentation:
- Every text model works through other API formats. Each model is listed with its native endpoint (OpenAI Responses, Anthropic Messages, and so on), and the console translates between them. That is how you can call a model from a client that speaks a different protocol.
- Auto-recharge and budgets. Auto-recharge buys credits when your balance runs low, and budgets cap monthly spend for the workspace and each member.
- Hosting and privacy. OpenCode says all Zen models are hosted in the US and EU, and its providers follow a zero-retention policy and do not use your data for training, with listed exceptions. OpenAI and Anthropic requests are retained for 30 days under those companies' data policies.
The free models: what you give up
The Zen catalog includes models with a -free suffix. OpenCode describes them as free for a limited time, and the team uses the period to collect feedback. Our catalog counted 11 such models at the last check, led by GLM 5.3 Flash (Artificial Analysis index 41.8) and DeepSeek V4.1 Flash (39.5). Crucially, the data terms are not the same for each, so read them:
| Free model | What OpenCode's docs say |
|---|---|
| GLM-5.3-Flash Free | Zero retention, no training on your data |
| DeepSeek V4.1 Flash Free | Zero retention, no training on your data |
| MiniMax M3.1 Flash Free | Zero retention, no training on your data |
| MiMo-V2.6-Flash Free | Collected data may be used to improve the model |
| Nemotron 3 Ultra, 3.5 Lightning, 3.5 Super VL (NVIDIA free endpoints) | Trial use only; do not submit personal or confidential data; usage is logged to improve NVIDIA products |
| Iceman Free (stealth) | Data is retained and used to improve the model |
| Flint 40B Free | Data may be used to improve the model; retained for 30 days |
| Inkling 2 Small Free | Data retained and may be used for training; not available in the EU, Iceland, Liechtenstein, Norway, Cuba, Iran, North Korea, Russia, Ukraine or Venezuela |
| Beam Free | No training on prompts; retained 30 days for safety; unavailable in a list of countries and regions that the docs enumerate |
| LongCat 2.5 Free | Hosted in China; prompts not used for training or retained; limit of 100 requests per day |
The rule of thumb: use the zero-retention free models for real code, and treat the stealth and trial models as public. If you work with proprietary repositories, hide the others entirely. OpenCode documents an environment variable for that, OPENCODE_DISABLE_FREE_MODELS=1, which hides free models in the client.
Free routes also carry a restriction that is easy to miss: OpenCode notes that free-tier chat models require OpenCode itself. To use paid models, or Go, from other clients you use the regular API endpoints.
Go: a subscription with dollar allowances
Go gives you "reliable access to popular open coding models" for a flat fee. The plans:
| Plan | Price | Notes |
|---|---|---|
| Go | $10 per month | Lower-cost access to the model list |
| Go Plus | $40 per month | Higher usage limits on the same models |
| Go Teams | $20 per member per month | Each member gets their own limits, $5 per month higher than Go on every model |
Go and Go Plus are for single-member workspaces, such as the Personal workspace created at sign-up. Either can be upgraded to Go Teams at any time, with the unused part of the month's allowance credited. A member who joins a team workspace mid-period pays $20 and receives a full monthly allowance.
How the limits work
Go measures usage in dollars. Each model has a monthly dollar limit, and three windows apply to it: a five-hour window at 20% of the monthly limit, a weekly window at 50%, and the monthly window at 100%. Cheaper models stretch further because they consume the allowance more slowly, and every response reports how much the request counted in an opencode.allowance field. The docs list token prices per model, the same for Go, Go Plus, and Go Teams. Examples from the table: GLM-5.3-Flash is $0.15 input and $0.50 output per 1M tokens, and DeepSeek V4.1 Flash is $0.15 and $0.60 off-peak.
Beyond the limits you can enable Extra Usage to pay for overage, or fall back to free models or a separate Zen balance.
The model list
The current Go list includes GLM-5.3-Flash, GLM-5.3, and GLM-5.2; Kimi K3.1 Pro, K3, and K2.7 Code; Qwen3.8 Max and Flash and Qwen3.7 Plus; DeepSeek V4.1 Flash and V4 Pro; MiniMax M3 and M3.1 Flash; MiMo models; LongCat; Step 5 Preview; Hy models; Grok 4.7; Claude Haiku 5.5; Gemini 4 Argon; and GPT 6 Luna and GPT 5.6 Luna. OpenCode notes that the list may change as it tests new models. Some entries are limited by region or by privacy settings: by default Go blocks paid endpoints that train on request data and allows Europe and the United States, so DeepSeek V4.1 Flash and V4 Pro require an owner or admin to select Global in Settings, Privacy.
Clients that work with Go
Go is designed for OpenCode and for coding agents that produce similar requests, and traffic is monitored for abuse. OpenCode asks clients to identify with their own user agent and send a stable session ID in the x-opencode-session header so routing and prompt caching work well. The docs list validated clients, including Hermes, Claude Code, Codex, ZCode, Pi, jcode, and the Kilo Code CLI, often with a minimum version. See our Hermes Agent guide for that integration.
Go versus other coding subscriptions
Go's $10 price is the lowest flat plan in our subscription comparison, but it buys open models rather than the frontier closed models. If you need Claude Opus or GPT-6-class models, compare it with the plans in our November subscription roundup. Many developers use both: a flat plan for everyday work and pay-as-you-go credits for hard problems.
Which should you pick?
| Your situation | Choice |
|---|---|
| Trying OpenCode, no budget | Free Zen models, preferably the zero-retention ones |
| Private repository | Paid Zen models or Go; disable free models |
| Daily agent work, predictable bill | Go or Go Plus |
| Occasional frontier-model use | Zen credits with a budget cap |
| Team | Go Teams, $20 per member |
FAQ
Is OpenCode Zen free?
Models with the -free suffix are free for a limited time. Everything else is billed per request from your credit balance.
What is the difference between Zen and Go?
Zen is pay as you go with a wide model list. Go is a $10, $40, or $20-per-member monthly subscription for a curated list of open coding models with dollar-based allowances.
Do free OpenCode models train on my code?
Some do and some do not. GLM-5.3-Flash, DeepSeek V4.1 Flash, and MiniMax M3.1 Flash free routes are documented as zero retention with no training. Others, including stealth and trial models, may use your data.
Can I use Go outside OpenCode?
Yes, with validated clients such as Claude Code, Codex, and Hermes, provided they send the session information OpenCode expects and a recognizable user agent.
How do I hide free models?
Set OPENCODE_DISABLE_FREE_MODELS=1, as described in OpenCode's documentation.
Where can I see the current free list?
On our OpenCode page, which reads the public Zen catalog on every check.
