Ox Alpha: An Anonymous 1M-Context Coding Model Is Free for a Week on OpenRouter and OpenCode

A stealth model called Ox Alpha appeared on OpenRouter and OpenCode on August 20 with no named developer, a 1,048,576-token context window, and a price of $0 for input and output. OpenRouter describes it as "a reasoning model designed for coding, sustained agentic work, and production workloads." OpenCode says the free window runs about a week and that usage is "near unlimited." (OpenRouter on X; OpenRouter model page; OpenCode on X)

The catch is the one that comes with every stealth model: you do not know who built it, the window closes when the provider says so, and the two hosts do not describe its data handling the same way. Those three facts decide what it is safe to run through it before the week is out.

Key Takeaways

• Ox Alpha is an anonymous stealth model released August 20, 2026, routed by OpenRouter as stealth/ox-alpha and served by OpenCode as Ox Alpha Free (x-preview-f-free). (OpenRouter model page; OpenCode Zen docs)
• 1,048,576-token context, up to 131,072 completion tokens, text plus image plus video input, text output, function calling, and JSON output without schema enforcement. (OpenRouter model page)
• Free on both hosts. OpenCode's August 20 post says free for the next week; its August 21 post says near unlimited and free on OpenCode Go for the next 6 days, not counted against Go usage. (OpenCode on X, Aug 20; OpenCode on X, Aug 21)
• OpenCode says it has capacity for 100 trillion tokens per day during the window. (OpenCode on X, Aug 20)
• Data handling diverges: OpenRouter says prompts and completions are retained by the provider and not used for training; OpenCode says zero retention and no training. (OpenRouter model page; OpenCode Zen docs)
• Nobody official has said who built it. Benchmark numbers and lab attributions circulating on social media are community claims, not vendor data.

What OpenRouter and OpenCode Actually Announced

OpenRouter posted the model at 9:02 PM UTC on August 20 as a "New stealth model" and a "frontier model built for efficient coding, sustained agentic work, and real-world production use," listing the 1M-token context and text, image, and video input. Its model page adds that the model suits "long-horizon software engineering, complex reasoning, and workflows that combine text with visual context." (OpenRouter on X; OpenRouter model page)

OpenCode's post landed three minutes earlier and was the louder of the two. It promised the model "free for the next week" with 1M context, multi-modal input, zero data retention, "generous rate limits, near unlimited usage," and "capacity for 100T tokens per day." By the time of writing it had passed 6.8 million views. The follow-up on August 21 extended the offer to OpenCode Go subscribers: "For the next 6 days, usage is near unlimited and completely free" and "won't count against your Go usage." (OpenCode on X, Aug 20; OpenCode on X, Aug 21)

Neither post names the lab. OpenRouter's model page is explicit that the model "is developed and operated by a third-party provider who has chosen to remain anonymous during this preview" and that OpenRouter "routes requests to it and is not its developer, owner, or provider." (OpenRouter model page)

The Spec Sheet

Everything below is from the OpenRouter model page and OpenCode's Zen docs. Nothing here comes from the model's developer, because the developer has not published anything.

(OpenRouter model page; OpenCode Zen docs; OpenCode Go)

The 131,072-token completion ceiling is the number worth noticing for agent work. It means a single turn can emit a large multi-file change or a long reasoning trace without truncation, which matters more in practice than the headline context figure. For comparison, the 1M-context rollout of GPT-5.6 Sol in Codex was the last time a million-token window reached a coding agent most readers already use.

The Data-Handling Wording Does Not Match

The two hosts describe what happens to your prompts differently, and the difference is not cosmetic. OpenRouter's page says: "Prompts and completions for this model are retained by the provider and are not used for training; all other use is governed by the Stealth Model Terms." OpenCode's docs say: "Its provider follows a zero-retention policy and does not use your data for model training." OpenCode's launch post lists "Zero Data Retention" as a bullet. (OpenRouter model page; OpenCode Zen docs; OpenCode on X, Aug 20)

Both statements agree that the provider does not train on your data. They disagree on whether the provider keeps it. One of two things is true: the provider has different terms with each host, or one host's wording is loose. Until the provider identifies itself and publishes its own terms, the safe reading is OpenRouter's: assume retention, assume no training. Do not route proprietary source, customer data, or credentials through a model whose operator you cannot name.

OpenCode makes the same distinction itself between its two stealth offerings. Its docs say Big Pickle, the other free stealth model on Zen, is collecting feedback during its free period, while Ox Alpha Free is the one described with a zero-retention policy. (OpenCode Zen docs)

Who Built It

Nobody has said. OpenRouter's stealth program exists precisely so a lab can gather production traffic on an unreleased model without attaching its name, and the model page's FAQ answers "who made it" with the anonymity statement rather than a vendor. (OpenRouter model page)

Social media is full of attributions and benchmark tables for Ox Alpha. None of them comes from the developer or from either host, so this article does not repeat them. The only numbers with a source are the ones in the table above. If the provider unmasks the model, this page will be updated with the vendor's own figures.

How to Try It This Week

On OpenRouter, the model is stealth/ox-alpha and works through the standard chat completions API and the playground. Function calling and response_format JSON are supported, so an existing agent harness that targets OpenRouter can point at it by changing the model string. (OpenRouter model page)

On OpenCode, the model is listed as Ox Alpha Free with the id x-preview-f-free, configured as opencode/x-preview-f-free and served from https://opencode.ai/zen/v1/chat/completions through the OpenAI-compatible SDK. Zen's pricing table shows free input, output, and cached reads. (OpenCode Zen docs)

On OpenCode Go, the $5-first-month, then $10-per-month subscription, the usage-limits table shows Ox Alpha Free as unlimited requests per 5-hour window for a limited time, where every other Go model carries a numbered allowance. The August 21 post confirms Ox Alpha usage does not count against Go usage during the window. Go works with OpenCode or, per its own page, "any agent," so the model is reachable from terminals that wrap OpenCode such as Warp. (OpenCode Go; OpenCode on X, Aug 21)

A sensible way to spend the week: run it against a task you have already solved with a model you pay for, on code you do not mind a stranger retaining, and keep the traces. OpenRouter's post asks users to "share feedback to improve the model," which is the whole point of a stealth release: the provider is buying evaluation data with free tokens. (OpenRouter on X)

What to Watch

Three things will settle whether Ox Alpha matters beyond this week. First, whether the provider unmasks it and at what price; free stealth previews end, and the follow-on price is what decides whether the 1M window is usable day to day. Second, whether the retention wording converges once the vendor publishes its own terms. Third, whether it lands in the coding tools readers already use, the way Grok 4.6 shipped straight into Cursor and OpenRouter on the same day. An anonymous model on a router is an experiment; a named model in your editor is a decision.

Sources

• OpenRouter on X, "New stealth model: Ox Alpha": https://x.com/OpenRouter/status/2090544970923184269
• OpenRouter, Ox Alpha model page: https://openrouter.ai/stealth/ox-alpha
• OpenRouter, Stealth provider page: https://openrouter.ai/provider/stealth
• OpenCode on X, August 20, "Ox Alpha (stealth model) is free for the next week": https://x.com/opencode/status/2090544355824038300
• OpenCode on X, August 21, "Ox Alpha is now available on OpenCode Go too": https://x.com/opencode/status/2090758645499728234
• OpenCode, Zen docs (model list, pricing, stealth notes): https://opencode.ai/docs/zen/
• OpenCode Go: https://opencode.ai/go