Ox Alpha: The Free Stealth Coding Model (and How to Run It in MeshCode)

On August 20, 2026, a frontier-class model showed up on OpenRouter with no lab behind it, no announcement, no system card, and a price of exactly $0 — for input and output alike. It calls itself ox-alpha, it's still free as of this writing, and it might be the most interesting thing to happen to coding agents all month.

Here's what's actually known about it, what's still rumor, and how to run it in MeshCode in about a minute.

What ox-alpha actually is

Strip away the mystery and the published facts are straightforward:

• 1M token context window (~1,048,576), with up to 131,072 completion tokens. That's whole-repository scale — you can point it at a large codebase's worth of material in one sitting.
• A reasoning model built for coding and sustained agentic work — OpenRouter's own description is "designed for coding, sustained agentic work, and production workloads," which is agent-speak for "long sessions, real tool calls, real code."
• Multimodal input — text, images, and video in; text out. Tool/function calling works; JSON output is supported but not schema-enforced.
• Free — OpenRouter's FAQ states it plainly: prompt and completion tokens are not charged during the preview. Throughput sits around 21 tok/s with 100% uptime since launch.
• Your prompts are retained by the provider — but per the published terms, not used for training.

That last pair is worth reading twice. Free, retained, not trained: great for experimenting and fan-out work, something to think about before you paste production secrets into it.

How good is it, honestly?

A viral claim early on said ox-alpha scored above 80% on DeepSWE, beating Claude Fable 5 and GPT-5.6 Sol. That number came from a 10-task sample — and its own author flagged the variance. Full 113-task runs land in the 58–63% band: alongside DeepSeek V4 Pro, and below the ~73% that GPT-5.6 Sol posts at max reasoning. Notably, roughly 10% of ox-alpha's benchmark losses were tool-call format failures, not reasoning failures — an agent-plumbing problem, which tends to improve fast.

That's still a strong result for a model that costs nothing. It's not a mysterious superintelligence; it's a credible frontier-adjacent coder that happens to be free while its authors stay anonymous. For agentic coding — where you often want many long, parallel sessions — "solidly mid-frontier and $0" is a bigger deal than "marginally better than the best."

Who built it? Nobody knows

The attribution hunt is half the fun. Names floated include Google, Xiaomi, DeepSeek, and Microsoft MAI — but the strongest theory, based on tokenizer fingerprinting that matched across every normalized test, points to Zhipu AI / Z.ai's GLM-5.3 family. Z.ai has previewed models anonymously before (the "Pony Alpha" episode), so the playbook fits.

One caveat if that theory is right: Zhipu AI is on the U.S. Commerce Department's Entity List, which is why some teams think twice about where their code and prompts flow. None of this is confirmed — the developer "has chosen to remain anonymous during this preview," and that's the only official word. Treat the attribution talk as what it is: informed speculation.

The catch: free previews end

There's no announced cutoff. Reporting around the launch estimated the free window could close anywhere from Aug 24–27 — and as of publishing, it's still $0. But "no announced end date" is not "no end date." If you want to evaluate ox-alpha, the time to do it is now, while every token is on the house.

One more piece of fine print: OpenRouter's free tier is rate-limited — 20 requests/minute, and 50 requests/day on an account without credits (1,000/day once you've bought $10+ in credits). For agent work that pushes you toward fewer, longer sessions — which is exactly what MeshCode panes reward.

How to run ox-alpha in MeshCode

MeshCode ships an OpenRouter integration — 418 models one toggle away — so ox-alpha is a first-class citizen, not a hack. Setup:

1. Open Settings → Models and flip OpenRouter on.
2. Paste your OpenRouter API key and hit Use OpenRouter (a free OpenRouter account is enough — ox-alpha costs nothing, so no credits are consumed).
3. Pick stealth/ox-alpha in the model picker. It appears in the composer badge at the bottom of the window — that badge is the model every pane now answers to.
4. Describe the work. Every pane streams the agent's thinking, tool calls, and command output live, with your context usage on the gauge.

Because ox-alpha carries a 1M context window, it's unusually well suited to MeshCode's whole premise: several agents, several projects, one window.

That's the real unlock. Running four frontier-class agents in parallel used to be a budgeting decision; on a free 1M-context model it's just... a Tuesday. Fan out one pane per microservice, a cleanup agent on your oldest repo, a migration agent on the fifth — and keep the premium models for the parts that genuinely need them.

Further reading

• Who made ox alpha? Every clue, ranked — tokenizer fingerprints, the Pony Alpha precedent, and the theories that don't hold up.
• Ox alpha benchmarks, explained — the 80% viral claim vs. what full runs actually measured.
• Free coding models on OpenRouter, August 2026 — ox-alpha isn't the only $0 option, and the rate limits matter.

Try it

Stealth models this cheap don't come along often, and they don't stay free forever. Download MeshCode, flip on OpenRouter, pick stealth/ox-alpha, and put a million tokens of context to work across as many panes as you have projects.

ox-alpha is developed by an anonymous third party and distributed via OpenRouter; details above are drawn from OpenRouter's listings and independent community reporting as of August 26, 2026, and may change as the preview evolves or attribution is confirmed. MeshCode is not affiliated with or endorsed by the model's developer or OpenRouter.