Ox Alpha Free AI Model: Setup, Privacy and Team Buyer Guide
Ox Alpha is free to try, has a 1M-token context window and is built for coding agents, but the access route changes the privacy answer. OpenCode Zen lists zero retention. OpenRouter says its provider retains prompts and completions, while its general stealth-model agreement grants broader rights. Treat the model as a temporary evaluation route, not a default destination for private source code.
The official display name is Ox Alpha, with the letter O. Many people search for 0x Alpha or 0xAlpha. This guide owns the model-specific setup, privacy and pilot decision. For tool economics, read our Claude Code versus OpenCode team-cost comparison. For portfolio architecture, use the multi-model coding-agent buying guide.
Need to test Ox Alpha without exposing a production repository?
Design a Safe AI PilotOx Alpha at a glance
| Question | Verified answer on 23 August 2026 | Buyer implication |
|---|---|---|
| What is it? | An anonymous third-party reasoning model for coding, sustained agent work and production workloads. | The operator and long-term support owner are undisclosed. |
| Is it free? | Prompt and completion tokens are currently priced at $0 through the listed preview routes. | Free is a current preview price, not a durable budget assumption. |
| Context and output | 1,048,576-token context and up to 131,072 completion tokens. | Large limits create headroom, but do not prove retrieval quality across a whole repository. |
| Inputs | Text, images and video in; text out. | It can inspect screenshots and visual context, subject to route and harness support. |
| Agent features | Tool calling, tool choice and structured output are supported. JSON Schema enforcement is not listed. | Validate tool arguments and structured responses in code. |
| Can you self-host it? | No public weights or identified developer are listed. | You depend on a hosted preview and cannot pin the underlying weights. |
The current official OpenRouter model record supports the specifications above and dates the release to 20 August 2026. It does not identify the developer. Claims that Ox Alpha is GLM, MiMo or another unreleased family remain speculation unless the provider discloses its identity.
OpenCode Zen or OpenRouter: which Ox Alpha route should you use?
| Route | Model ID | Current token price | Published data position | Best fit |
|---|---|---|---|---|
| OpenCode Zen | x-preview-f-free, or opencode/x-preview-f-free in OpenCode config | Free | Provider follows zero retention and does not train on your data. | A redacted coding-agent evaluation where this policy is acceptable. |
| OpenRouter | stealth/ox-alpha | Free | Model page says the provider retains prompts and completions but does not use them for training. | API experimentation with no confidential or personal data. |
The OpenCode Zen documentation confirms the ID, zero token price, limited-time status and route-specific zero-retention statement. It also documents workspace and member spending limits. Those controls help with paid models and add-ons, but they do not turn temporary model availability into an SLA.
Is Ox Alpha private enough for company code?
Not by default. A repository can contain credentials, customer identifiers, private issue text, proprietary algorithms and licensed code. A zero-dollar model invoice does not transfer permission to disclose those assets.
There is also a material policy difference inside OpenRouter. The model page says Ox Alpha content is retained by the provider and not used for training. The broader OpenRouter Stealth Program agreement says user content may be collected, retained and provided to anonymous providers for model training and improvement. It also says stealth models can disappear at any time and prohibits sensitive categories of data.
When a product page and a general agreement appear to differ, a cautious buyer should follow the stricter boundary until counsel or the platform confirms the applicable terms in writing. For OpenRouter, do not send private repositories, personal data, health data, financial data or customer material. For OpenCode Zen, verify the current policy again at pilot start, record it in the evaluation, and still remove secrets. Our PII-redaction guide for LLM prompts provides the preprocessing pattern.
How to use Ox Alpha Free in OpenCode
- Choose the route first. Decide whether the OpenCode Zen or OpenRouter terms fit the data classification of your test.
- Create a separate pilot key. Give it the lowest practical budget and no access to unrelated environments.
- Connect the provider. In OpenCode, run
/connect, choose OpenCode Zen or OpenRouter, and paste the matching API key. - Select the exact model. Run
/models. Choose Ox Alpha Free on Zen orstealth/ox-alphaon OpenRouter. Record the model ID in every result. - Start with a sanitized repository. Use a copy with secrets removed, synthetic fixtures and no customer data.
- Constrain permissions. Require approval for network access, package installation, destructive shell commands and writes outside the worktree.
- Verify independently. Run deterministic tests and have a separate reviewer assess correctness and security.
The official OpenCode provider setup guide documents the /connect and /models flow for both Zen and OpenRouter. Do not paste API keys into project configuration or commit them.
Is Ox Alpha really free, or are there hidden costs?
The model tokens are currently free. The complete workflow might not be. A coding harness can invoke search, file parsing, external MCP tools, storage and other paid services. It can also generate engineering cost through retries, slow responses, failed tasks and human review.
| Cost layer | What can happen | How to control it |
|---|---|---|
| Model input and output | Currently $0 on the documented preview routes. | Recheck pricing before every evaluation batch. |
| PDF and file processing | A gateway may parse a file through a separate paid engine. | Select the parser explicitly and inspect usage logs. |
| Web search and tools | External tool calls can have their own metering. | Disable unneeded tools and set independent limits. |
| Rate limits and outages | Free previews can be slower, constrained or removed. | Use bounded retries and a named fallback model. |
| Human review | Large outputs can increase review time and defect surface. | Ask for small diffs and measure correction minutes. |
For example, the OpenRouter PDF-input documentation lists a paid Mistral OCR option alongside free Cloudflare parsing and native processing. If no engine is selected, routing can choose a parser for you. A zero-cost model can therefore coexist with a non-zero request.
What is Ox Alpha good for?
- Long-repository exploration: compare its architecture map and cited file paths against a trusted baseline.
- Multi-file implementation: test whether it keeps requirements, tests and local instructions aligned across a bounded change.
- Visual bug triage: combine a sanitized screenshot with code and a reproducible error report.
- Low-cost evaluation: build an internal task set without spending model-token budget during the preview.
A 1M window is capacity, not evidence of perfect recall. Test buried constraints at different positions, conflicting instructions, tool recovery and completion quality. If your real question is whether long context should replace retrieval or adaptation, use our RAG versus fine-tuning versus long-context decision guide.
Who should not deploy Ox Alpha?
- Teams that need a named model developer, fixed weights, contractual support or a long-term availability commitment.
- Workloads containing secrets, regulated data, customer repositories or data you cannot lawfully share with the selected route.
- Production automation without strict tool permissions, output validation, rollback and a fallback model.
- Buyers choosing from screenshots, self-reported identity answers or unrepeatable benchmark claims.
A seven-day Ox Alpha evaluation plan
- Day 1, contract and data gate: archive the current route terms, classify the test data and remove secrets.
- Day 2, baseline: freeze 15 to 25 representative tasks, repository commits and acceptance tests.
- Days 3 and 4, repeated runs: run Ox Alpha and one approved fallback at least three times per task class.
- Day 5, adversarial review: test denied files, prompt injection, tool errors, stale documentation and ambiguous requirements.
- Day 6, commercial scorecard: calculate accepted changes, serious defects, latency, reviewer minutes and total non-model charges.
- Day 7, decision: approve a narrow use, keep it in experimentation, or reject it. Define the exit trigger if pricing, identity, terms or model behavior changes.
Recommendation: use Ox Alpha as a disposable evaluation lane for sanitized coding tasks. Do not build a production dependency around an anonymous, time-limited preview. If the pilot succeeds, ask Wavect to turn the results into a governed AI-agent rollout with routing, evaluation, observability and fallback ownership.
Sources and research boundary
This guide was researched on 23 August 2026 from the current OpenRouter model page and stealth terms, plus OpenCode's Zen and provider documentation. Pricing, retention statements, model IDs and availability can change without preserving this article's conclusions. Recheck the linked primary sources before sending data or approving a rollout. We intentionally do not identify the anonymous developer and do not repeat unverified benchmark scores.
Frequently Asked Questions
Is the model called Ox Alpha or 0x Alpha?
Is Ox Alpha free?
Does Ox Alpha have a 1M context window?
Is Ox Alpha a GLM or MiMo model?
Does Ox Alpha retain code or prompts?
Can a team use Ox Alpha in production?
Final thoughts
Ox Alpha is an unusually capable free preview on paper: 1M context, large outputs, multimodal input and agent features. The durable buying answer is more cautious. Its provider is anonymous, access is temporary, and the published data position changes by route.
Choose the route before the model. Remove secrets, constrain tools, evaluate on fixed tasks and price every accepted change, including add-ons and review. Keep a fallback. Promote only a narrow use case after the evidence and terms meet your team's bar.
