Trending on Kingy
Keep reading with the stories getting the most attention now.
OpenCode has put an anonymous AI coding model called Ox Alpha in front of developers at zero token cost for one week. The offer is worth trying if you have a disposable branch, a difficult repository task and time to inspect every change. The model’s claimed 1M-token context window, multimodal inputs and unusually generous capacity make this a rare chance to probe a potentially serious coding agent without watching a usage meter.
Treat the week as a public stress test, not a production migration. Nobody has publicly identified the developer. Independent benchmark evidence is thin, post-preview pricing is unknown, and OpenCode has already advised users with network errors to update the client. The free window is an invitation to test hard, but the model has not earned blind trust.
The answer in one sentence: Ox Alpha is worth a focused test this week for developers who can verify its work in a safe repository, provided they keep secrets out of prompts and separate OpenCode’s privacy terms from other access routes.
Quick verdict
| Best for | Skip if |
|---|---|
| Whole-repository mapping, multi-file refactors, stubborn bug hunts and multimodal coding experiments | You need a named provider, stable production support, known pricing or audited security assurances |
| Developers willing to run tests, inspect diffs and retry interrupted work | You plan to paste confidential source code into whichever free endpoint appears first |
| Learning how a long-context model behaves inside an agent client | You need independent proof that the full 1M-token window remains useful at maximum length |
Promotional price checked: August 21, 2026. OpenCode lists Ox Alpha inference as free for a limited time. OpenCode Go itself costs $5 for the first month and $10 per month after that. OpenCode Zen asks users to add billing details during setup.
Testing scope: Kingy verified the live announcement, OpenCode documentation, current route names, pricing language and privacy disclosures. We did not run Ox Alpha because this test environment had no OpenCode client or configured OpenCode/OpenRouter API access, and we did not create a paid account or add billing details. Performance observations below are therefore documentation-based or explicitly attributed to community reports.
What Ox Alpha is, and what remains hidden
OpenCode’s August 20 announcement describes Ox Alpha as a stealth model and says it will be free for the next week. OpenCode advertises a 1M-token context window, multimodal support, zero data retention, generous rate limits and capacity for 100 trillion tokens per day.
The anonymous label matters. No public source currently names the lab, underlying model family, architecture, training data, license or release plan. Ox Alpha could eventually appear under a familiar product name. It could also vanish when the preview ends. Any confident attribution to GLM, MiMo, Qwen, Gemini or another family is still guesswork.
OpenRouter’s current Ox Alpha listing adds useful detail. It describes a reasoning model for coding and sustained agent work with a 1,048,576-token context window, a 131,072-token maximum output, text/image/video inputs and text output. The route is stealth/ox-alpha and is currently priced at $0 per million input and output tokens.
Those specifications describe capacity, not proven quality. A model can accept a huge prompt and still lose details or choose the wrong files. The practical test is whether it can navigate a real codebase, preserve constraints across tool calls and finish with working tests.
Ox Alpha facts: confirmed, claimed and unknown
| Item | Current information | Evidence status |
|---|---|---|
| Developer | Undisclosed third-party provider | Confirmed undisclosed by the stealth listings |
| OpenCode availability | Zen and Go | Confirmed in OpenCode documentation |
| Promotional price | $0 token price for a limited time | Confirmed; account or subscription friction depends on route |
| Offer duration | “For the next week” from August 20 | OpenCode announcement; no precise cutoff time published |
| Context window | 1,048,576 tokens | Provider/OpenRouter specification; not independently stress-tested by Kingy |
| Maximum output | 131,072 tokens | OpenRouter specification |
| Inputs | Text, images and video | Provider/OpenRouter specification |
| OpenCode data retention | 0 days | OpenCode policy table |
| OpenCode training use | Not used | OpenCode policy table |
| Rate limits | “Generous” and “near unlimited” | OpenCode claim; limits may change |
| Daily capacity | 100T tokens | OpenCode claim; not an individual allowance |
| Post-preview price | Unknown | No published figure found on August 21 |
Free has three different meanings here
The cleanest way to understand the offer is to separate the routes.
| Route | Current model ID | Cost and setup | Data handling disclosed for Ox Alpha |
|---|---|---|---|
| OpenCode Zen | opencode/x-preview-f-free |
Ox Alpha tokens are free; Zen setup asks for billing details and an API key | OpenCode says zero retention and no model training |
| OpenCode Go | opencode-go/ox-alpha-free |
Go costs $5 for month one, then $10/month; Ox Alpha is listed as free for a limited time | OpenCode lists 0-day retention and no model training |
| OpenRouter | stealth/ox-alpha |
Currently $0 input and output | The provider may retain prompts and completions but says it does not use them for training |
OpenCode Zen’s pricing table lists Ox Alpha Free at $0 for input, output and cached reads. The same page says users sign in, add billing details and copy an API key. Its model table currently maps the public display name to x-preview-f-free, so readers following an older social post may not see “ox-alpha” in their configuration.
OpenCode Go uses the clearer ox-alpha-free ID, but Go is a subscription. Its documentation says free models remain usable after the normal plan limits are reached, while warning that limits can change as OpenCode learns from early usage.
A zero token price does not guarantee zero onboarding cost, and a privacy statement attached to one gateway should not be copied onto another. Check the route before sending code.
How to select Ox Alpha in OpenCode
OpenCode documents the same basic connection flow for Zen and Go:
- Sign in to OpenCode Zen. For the Go route, subscribe to Go and copy the Go API key. For Zen, follow the billing and API-key setup shown in the console.
- Start OpenCode in the project you want to test.
- Run
/connectin the terminal interface. - Choose OpenCode Zen or OpenCode Go, then paste the matching API key.
- Run
/modelsand select Ox Alpha Free. - Confirm the displayed route. As of August 21, Zen documents
opencode/x-preview-f-free, while Go documentsopencode-go/ox-alpha-free.
Use the selector instead of hard-coding an ID copied from a screenshot. The aliases can change during a stealth preview.
Reliability is part of the test
The launch generated reports of sessions stopping mid-task, provider errors and rate-limit messages. Anecdotes cannot establish a failure rate, but OpenCode’s response makes the issue material: the company advised Ox Alpha users with network errors to upgrade to OpenCode 1.18.21, which retries unknown stops automatically.
That update may make interrupted runs easier to recover, but it cannot prove that the endpoint is stable. Record the OpenCode version, route and exact task. If a run stops, distinguish a client retry from a model failure or provider outage.
The 100T-per-day figure deserves similar care. OpenCode presented it as system capacity, not a personal quota or service-level agreement. “Near unlimited” should be read as promotional capacity during a preview. It is not a promise that every request will start immediately or finish successfully.
Five tests worth running before the window closes
The free period is most valuable when the task produces an objective result.
- Map a repository. Ask for entry points, data flow, state boundaries, test strategy and the five files most likely to change for a feature. Check every path.
- Attempt a cross-file refactor. Use a disposable branch, define the expected behavior and require the existing tests to stay green. Check callers, types, fixtures and documentation.
- Repair a seeded bug. Introduce a known defect behind a failing test. Supply the failure output, but not the answer. Score diagnosis, patch quality and regression risk.
- Try screenshot-to-code work. Provide a synthetic interface and component library. Evaluate layout accuracy, responsiveness and invented assets.
- Run a long agent task. Choose work that needs planning, tool use and verification. Record stops, retries, malformed tool calls and incomplete work.
Keep the branch disposable. Inspect the diff before running generated code, and run tests in a constrained environment. A free model can still produce an expensive mistake.
Privacy: useful policy, limited assurance
OpenCode’s Zen documentation says the Ox Alpha provider follows a zero-retention policy and does not use customer data for model training. The Go privacy table lists zero days of retention and no training use for Ox Alpha Free.
That policy is not an independent security audit, a public data-processing agreement or proof of the provider’s identity. Regulated and client-confidential work needs stronger assurance.
OpenRouter currently discloses different handling for its Ox Alpha route: prompts and completions may be retained by the provider, although the provider says they are not used for training. Developers should not move between the OpenCode and OpenRouter endpoints while assuming the same privacy terms follow the model.
For this preview, use public repositories, synthetic projects or code you are authorized to disclose. Keep credentials, customer data, private incident logs and proprietary source out of every free endpoint unless your organization has approved the exact provider route.
The identity hunt is entertaining, not actionable
Developers are comparing tokenizer behavior, tool-call patterns, video token counts and writing style to existing model families. The leading theories change by the hour. None of those clues names the developer.
Behavioral similarity can narrow a guess, but inference stacks and post-training can distort the fingerprint. A matching tokenizer may suggest lineage or reuse; it does not prove ownership. Until the provider speaks, “unknown” is accurate.
The launch strategy is clearer than the model’s identity. A provider with substantial inference capacity wants developers to push long, multimodal agent tasks before the public name and price are fixed. The week functions as a public audition.
Who should try Ox Alpha
Ox Alpha makes sense for developers who already know how to evaluate an agent: small reversible commits, explicit acceptance tests, restricted permissions and a willingness to inspect the result. It is especially interesting for teams that have been priced out of long-context experiments or want to see how a model handles a repository too large for a conventional coding chat.
Production teams should wait if model provenance, uptime commitments, support ownership or stable pricing are mandatory. Security-sensitive organizations should wait for contractual terms tied to a named provider. New developers can still experiment, but they should start with a toy repository rather than an important project.
Final verdict
Ox Alpha has earned attention, not trust. The free week removes the token bill from a class of tests that can become expensive quickly, and its published context and multimodal limits are unusually ambitious. The unknown provider, route-specific privacy terms and early reliability friction keep the model in experimental territory.
Try it on one demanding, measurable task before the offer disappears. Use a safe repository, record the route and client version, inspect every change and let the tests decide whether the mystery model deserves a place in your workflow.
Ox Alpha FAQ
Is Ox Alpha really free?
Its token price is currently zero on the listed preview routes. Access conditions differ. OpenCode Zen asks for billing details, OpenCode Go is a paid subscription, and OpenRouter currently lists the model at $0. Check the live terms before connecting.
When does the free Ox Alpha offer end?
OpenCode said on August 20, 2026 that the model would be free “for the next week.” It did not publish a precise cutoff time and timezone in the announcement, so an exact deadline would be an inference.
Who made Ox Alpha?
The provider has not been publicly identified. Claims that Ox Alpha is a specific unreleased GLM, MiMo, Qwen or other model remain speculation.
Does Ox Alpha retain my code?
OpenCode says its Ox Alpha routes use zero retention and do not use prompts for training. OpenRouter’s listing says its provider may retain prompts and completions but does not use them for training. The answer depends on the route.
Does Ox Alpha really support one million tokens?
The published context limit is 1,048,576 tokens. That establishes the accepted window, not reliable recall or reasoning across every token. Kingy did not independently test the full window.
Which model ID should I use in OpenCode?
As of August 21, OpenCode documents opencode/x-preview-f-free for Zen and opencode-go/ox-alpha-free for Go. Run /models and select the displayed Ox Alpha Free entry because preview aliases can change.
What should I do if Ox Alpha stops mid-task?
Update OpenCode to version 1.18.21 or later, then retry from a clear checkpoint. Record whether the failure came from the model, provider or client. Do not treat repeated retries as a completed task.