How to Use Ox Alpha in OpenCode: Free Setup

Quick answer: Ox Alpha Free is already live on OpenCode Go — a one-week, near-unlimited free window (the announcement cites 100T tokens/day capacity) that does not count against your Go usage. Open OpenCode, switch to the Go model list, pick Ox Alpha Free, and start an agentic coding session. No API key, no provider account, no configuration file required. If you'd rather use your own key, the same model is also reachable through OpenRouter as stealth/ox-alpha.

If you code agentically, you've felt the squeeze: every capable model either burns through your plan quota inside a few long sessions, requires a paid API key before you've written a line, or caps context so aggressively that multi-file tasks degrade into guesswork. A model that removes all three frictions at once — free, near-unlimited, 1M-token context — sounds too good to be true. That's exactly what the Ox Alpha Free listing on OpenCode Go is offering this week, so let's walk through what it actually is, how to turn it on, and the honest limits before you trust it with your codebase.

This guide is based on the OpenRouter official model listing, the OpenRouter public API data, and the official OpenCode announcement (checked August 21–22, 2026). We have not run our own benchmarks; where numbers come from community tests, they are labeled as such.

What the OpenCode Announcement Actually Says

Ox Alpha (listed as "Ox Alpha Free") appeared on the OpenCode Go model list on August 21, 2026, with a limited-time offer: free for the next week, with near-unlimited usage during the window. Two details in the announcement matter more than the headline:

It won't count against your Go usage. Sessions on Ox Alpha Free are separate from your Go plan quota, which means heavy agentic runs during this week don't eat into the allowance you pay for.
Zero data retention is claimed on the Go listing, alongside 1M context and multimodal input — the model accepts text, image, and video input and returns text.

The positioning matches the model page on OpenRouter: Ox Alpha is "a reasoning model designed for coding, sustained agentic work, and production workloads," suited for "long-horizon software engineering, complex reasoning, and workflows that combine text with visual context." It launched on OpenRouter on August 20, 2026 at $0 / $0 during the preview.

Step 1: Open OpenCode and Find Ox Alpha Free

No setup files are needed for the OpenCode Go path — that's the entire appeal:

Launch OpenCode from your terminal as you normally would (opencode).
Open the model picker (the same flow you use to switch between your configured models).
Look for the OpenCode Go model list — Ox Alpha Free appears there alongside the other Go-provided models.
Select Ox Alpha Free.

UI labels and menu positions vary between OpenCode versions, so don't memorize a screenshot: the model you're looking for is the Go-listed one named "Ox Alpha Free." If your OpenCode install doesn't surface Go models, make sure the client is on a recent release and is signed in with the account that has Go access.

Step 2: Start an Agentic Session

Once selected, just start coding. What you get under the hood:

1,048,576-token context (1M) — large enough for whole repositories or long multi-file change sets.
131,072-token maximum output per request — room for long reasoning traces plus substantial edits.
Mandatory reasoning with default effort max (the API also accepts high and low), so long-horizon tasks get full deliberation by default.
Tool calling and structured output — the API exposes tools, tool_choice, and response_format, which is exactly what an agentic coding loop needs.

OpenCode's published monitoring for the model (first ~3 days after launch) shows ~24 tokens/s throughput (P50) and ~5.81s latency (P50), with 99.99% uptime over 3 days. It's a preview model, so treat the speed numbers as directional.

Alternative Path: Connect Ox Alpha via an OpenRouter Key

Prefer to bring your own key? The same model is available through OpenRouter under the ID stealth/ox-alpha, using the standard OpenRouter-in-OpenCode flow:

Get an OpenRouter API key and make sure it's available to OpenCode as an environment variable.
Configure OpenRouter as a provider in your OpenCode config (model ID stealth/ox-alpha).
Optionally tune the reasoning effort (max / high / low) in your model settings.
Start a session and confirm the request routes to stealth/ox-alpha.

Because the model is in a free preview, prompt and completion pricing on OpenRouter are currently $0 / $0 — but remember that a free window, wherever you access it, can end without notice. OpenCode's own docs are the right place to confirm the current provider configuration syntax for your version.

The Free Window and Its Limits

Three honest caveats about "free":

The window is one week. The announcement says "free for the next week," and the OpenRouter page lists the preview as "PRICE: Free." What happens after the window — pricing, quota, or availability — is not disclosed.
Near-unlimited is not unlimited. The announcement cites 100T tokens/day capacity. That's an enormous ceiling, but it's a preview capacity figure, not a guarantee of unlimited throughput for every user at peak times.
It's a preview. The provider identity, architecture, and parameter count are not disclosed. Performance and behavior can change while the preview lasts.

Community tests circulating this week report roughly 80% on a DeepSWE 10-task subset (community-reported, not independently verified; same small sample: Fable 65%, GPT-5.6 Sol 52%) and a Kingbench 87.5% claim (community-reported, not independently verified). OpenRouter's model page itself publishes no benchmark scores, and independent tracker Artificial Analysis does not list the model (checked 2026-08-22). Treat all of these as directional signals, not a ranking.

What Ox Alpha Is Good For in OpenCode

The usage signal from OpenRouter's first two days is unusually consistent: the five largest apps by token volume are Hermes Agent (120B tokens), Claude Code (108B), Oh-My-Pi (93.5B), DeepSeek Harness multimodal-bridge (84.8B), and ZCode (51.5B) — every single one is an agentic coding or agent tool. Total volume hit roughly 657B prompt tokens and 7.95B completion tokens in that window. Real users are using it for agent work, so the right tasks are:

Repo-scale tasks in long context — the 1M window covers whole monorepos and long multi-file refactors without mid-task summarization.
Multi-step agent loops — tool calling plus mandatory reasoning fits OpenCode's edit-test-iterate cycles.
Visual-context coding workflows — text + image + video input means you can feed screenshots of UI bugs or video captures of broken flows directly into the loop, not just paste error text.

For quick single-line fixes, a small fast model is still the more sensible tool; Ox Alpha's value is in the long-horizon, context-heavy work. If you want to poke at the model before wiring it into a full agent loop, try Ox Alpha free on glm5.app — a browser chat that points at the same stealth/ox-alpha model, no terminal required.

How to Choose a Free Coding Model After the Window Ends

Here's the part most setup guides skip: the free week will end, and you'll need a decision framework, not another announcement. Score candidates on four axes:

Context size and how it degrades. A 1M-token window isn't just about capacity — it changes how much of your repo fits before the agent starts summarizing. Compare effective context after tool-call overhead, not the headline number.
Tool-calling maturity. Agentic coding lives or dies on tools/tool_choice reliability and structured output. Check the provider's API surface for these parameters and whether the model actually follows tool calls across long loops. Ox Alpha explicitly exposes tools, tool_choice, response_format, and reasoning_effort — that's the baseline to demand.
Data policy. What happens to your prompts and completions? The OpenCode Go listing claims zero data retention for Ox Alpha Free; on the OpenRouter side, the Stealth Terms state the provider retains prompts and completions but does not use them for training. Different access paths, different policies — read the one that applies to your setup before pushing proprietary code through any free model.
Real cost per completed task, not per token. $0/1M tokens is unbeatable per token but irrelevant if the model needs three attempts. The honest question is total cost (including your time) per merged feature. After the preview, pricing is not disclosed — so re-check the official page when the window ends rather than assuming today's $0 persists.

One more signal worth watching: Ox Alpha's top apps are all agent tools, which tells you what the community actually validated it for. When comparing models, look for the same kind of real-world usage evidence — not just marketing benchmarks.

Caveats Before You Trust It With Your Codebase

The provider is anonymous. Ox Alpha is a stealth model — developed and operated by a third-party provider that "has chosen to remain anonymous during this preview." OpenRouter routes requests and is not its developer, owner, or provider.
Data handling differs by access path. Via OpenRouter, prompts and completions are retained by the provider (not used for training, per the Stealth Model Terms). Via OpenCode Go, the listing claims zero data retention. Either way: don't paste API keys, secrets, or sensitive customer code into a preview session and assume it's private.
Everything can change. Pricing, quotas, availability, even the model's behavior are all provisional during the preview. If a free window matters for a production decision, treat the "free" as a trial, not a dependency.

FAQ

Is Ox Alpha free in OpenCode? Yes — Ox Alpha Free is on the OpenCode Go model list with a one-week free window and near-unlimited usage (100T tokens/day cited) that doesn't count against your Go usage.

How long is it free? The OpenCode announcement says "free for the next week" (announced August 21, 2026); OpenRouter's page lists the preview at $0 / $0. Post-window pricing is not disclosed.

Do I need an API key to use it in OpenCode? Not on the OpenCode Go path — just pick Ox Alpha Free from the Go model list. If you use the alternative path, you can connect it via OpenRouter with an OpenRouter key and model ID stealth/ox-alpha.

What tasks is Ox Alpha good for in OpenCode? Long-context, repo-scale agentic work: multi-file changes, multi-step agent loops, and workflows that combine visual context (image/video input) with code. Its top usage comes from agentic tools like Hermes Agent, Claude Code, and ZCode.

Does it count against my OpenCode Go usage? No — the announcement explicitly states the free window "won't count against your Go usage."

If this week's window has you curious about the model itself, start a free Ox Alpha chat on glm5.app to get hands-on before you commit any workflow to it.