OpenCode models: the free ones, Go's lineup, and switching

What opencode models shows

On OpenCode 2, opencode models prints every model you can use right now, one provider/model per line, such as opencode/big-pickle or opencode-go/kimi-k3. Its only flags are --server and --standalone; version 1’s provider filter, --refresh and --verbose are gone. The list grows as you connect providers, and it also changes by itself. OpenCode reads its catalogue from OpenCode’s models.dev database, refreshes it every five minutes and leaves out models marked deprecated. On a fresh install with no account, the ten free models listed at 12:37 UTC on October 22, 2026 had become a different ten by 12:41: LongCat 2.5 Preview Free had gone and Flint 40B Free had taken its place, with nothing updated on the machine.

OpenCode’s models docs listed fourteen free models that day. Three of them (LongCat 2.5 Preview Free, Exo Free and MiMo-V2.5 Free) were ones the catalogue had already retired, and one, Jev 1.13 Free, makes typed decisions rather than writing code. What you can actually use is what opencode models prints.

The free models, and their catches

In the docs’ words, “Free means an opencode or opencode-go model with no input price.” You need no account for them. Each one comes with conditions:

• They are temporary. Every one is described as free “for a limited time”, and no page gives an end date. In the month to October 22, OpenCode’s catalogue added thirteen free models and retired at least nine, one of them on the day it arrived.
• Some use your data. For Big Pickle, MiMo-V2.6-Flash Free and Flint 40B Free, the docs say “During its free period, collected data may be used to improve the model.” Flint 40B Free keeps it for 30 days. The three Nemotron models run on NVIDIA’s free endpoints: “Trial use only — do not submit personal or confidential data.” Beam Free’s provider doesn’t train on your data but keeps it for 30 days, and Beam Free isn’t offered in a list of countries and regions, China and Russia among them. DeepSeek V4.1 Flash Free, GLM-5.3-Flash Free and MiniMax M3.1 Flash Free are on zero-retention providers that don’t train on your data.
• There is a daily cap. When you reach it, OpenCode says: “Free usage limit reached. Connect a provider or subscribe to OpenCode Go to keep going, or switch models.” The Go page gave the cap as “200 requests/day” on October 22, 2026, a line OpenCode had already removed from its source for the next update. No other page gives a number.

For trying OpenCode out, the free models are fine. For code you wouldn’t paste into a public form, use a provider whose terms you have read. On a shared or team machine, setting OPENCODE_DISABLE_FREE_MODELS=1 on the server process hides them, and opencode service set env OPENCODE_DISABLE_FREE_MODELS 1 does it for the background server.

OpenCode Go’s models

Go is OpenCode’s subscription to a lineup it has tested for coding: Go at $10 a month, or Go Plus at $40 for higher limits on the same models. On October 22, 2026, the Go page and the Go docs listed the same 29 models. The docs call them “popular open coding models”, and the lineup also carries GPT 6 Luna, GPT 5.6 Luna, Claude Haiku 5.5, Grok 4.7 and Gemini 4 Argon Contributor from OpenAI, Anthropic, xAI and Google. Two kinds trade your data for the price: Gemini 4 Argon Contributor (“Your prompts and completions may be used to train future Google models”) and the two Muse Spark Contributor models, which do the same for Meta.

Each model has its own monthly allowance in dollars, the same token prices on both plans, and 20% of it usable in any 5 hours and 50% in a week. The docs estimate what that means in requests:

The docs’ 29 still included LongCat 2.5 Preview Free, which OpenCode’s catalogue had retired that morning and Go’s own model list left out; here too, what you can use is what opencode models prints. The estimates assume a typical Go request, and the docs say actual usage varies. At a model’s limit, you fall back to the free models, or to your pay-as-you-go balance if you turn on Extra usage. In OpenCode, run /connect, choose OpenCode Go and paste the key from the Console; Go’s models then appear as opencode-go/<model-id>. OpenCode’s pay-as-you-go catalogue, which its site calls Zen and its version 2 docs and the app call OpenCode Console, connects the same way and lists its models as opencode/<model-id>, the free ones among them. What Go and that catalogue cost beside your other options is in OpenCode’s pricing.

Choosing a model, and a default

In the terminal interface, /models or Ctrl+X then M opens the picker for the current session; it shows only enabled models whose provider is set up in this project, and a choice there applies to that session without changing your config. F2 steps through recent models, Ctrl+F marks a favourite, and Ctrl+T cycles a model’s variants, such as its reasoning effort (all from the keybind docs). In the desktop app and the web interface, Settings, then Models, chooses which models appear in the app’s model picker.

For one run, name it: opencode run -m opencode-go/kimi-k3 "Fix the failing test", with an optional #variant such as #high where the model offers one. For a default, set the root model key in ~/.config/opencode/opencode.json, or in a project’s own opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "opencode-go/kimi-k3"
}

Three details from the models docs:

• The root default keeps no variant; an agent’s own model can carry one.
• If the default isn’t available, OpenCode “falls back to the newest available supported model”.
• Model IDs are case-sensitive, and are the catalogue’s IDs, not display names.

A version 1 config keeps working: the migration page says “The V1 config format remains supported”, and version 2 reads the old fields as their new equivalents. provider becomes providers, small_model becomes the model of the built-in title agent, and enabled_providers and disabled_providers become policies. Moving to the new keys is optional.

Your own provider or key

/connect lists the providers your OpenCode can use and asks for whatever each needs: an API key, a browser or device sign-in, or a command. Outside the interface, opencode auth login does the same, and opencode auth login anthropic --method key goes straight to key entry. A provider’s usual environment variable, such as ANTHROPIC_API_KEY, also works when it is set on the server process. For the background server that means opencode service set env ANTHROPIC_API_KEY …, which stops it until the next command starts it again. Saved keys and sign-ins live in the server’s database, ~/.local/share/opencode/opencode.db. OpenRouter models take their full OpenRouter ID after the prefix, as in openrouter/anthropic/claude-sonnet-4. Which subscriptions OpenCode can use, and why a Claude plan isn’t one, is in OpenCode’s pricing.

A local model: Ollama, LM Studio, vLLM

OpenCode 2 finds local servers by itself at their default addresses, according to its providers docs. Ollama is found at 127.0.0.1:11434, LM Studio at 127.0.0.1:1234 and vLLM at 127.0.0.1:8000. Their models appear in /models with no account, as ollama/qwen3:8b or lmstudio/google/gemma-4-26b-a4b. Ollama’s embedding-only models are left out. vLLM’s models start with tools off, since OpenCode can’t tell whether the server was started with tool calling (--enable-auto-tool-choice). For Ollama on another host or port:

{
  "providers": {
    "ollama": {
      "settings": { "baseURL": "http://192.168.1.20:11434/v1" }
    }
  }
}

For any other OpenAI-compatible server, llama.cpp’s llama-server among them, the docs give a providers entry with the package @opencode/ai/providers/openai-compatible, a baseURL, and each model’s real name, context and output limits and tool support. OpenCode can’t detect those for an unlisted model, and assumes 200,000 tokens of context and tool support unless told otherwise.

Ollama’s own page offers a shortcut, ollama launch opencode, which starts OpenCode with the model you pick without touching your config, and it says “OpenCode requires a context length of 64k or higher.” Its manual setup still shows version 1’s provider key; on version 2 the key is providers.

Which models people use

OpenCode publishes no ranking of models by quality. What it does publish is OpenCode Data, usage across OpenCode by model, refreshed every hour. At its 1:01 PM UTC update on October 22, 2026, it read: “DeepSeek V4.1 Flash led OpenCode usage over the past 7 days with 103T tokens, followed by MiniMax-M3.1-Flash (43T) and Muse Spark 1.3 Contributor (11T).” The top of the list is led by models OpenCode offers free or in Go, so it shows what costs least to run in OpenCode as much as what works best.