Verdict: Ox Alpha is a genuinely useful, completely free coding model with a 1,048,576-token context window that you can use right now through OpenRouter or OpenCode, and it is good enough to take seriously in the best open-source LLM for coding conversation. The catch: it is an anonymous "stealth" preview, the provider retains your prompts, and the free window is temporary. Use it for non-sensitive code, side projects, and evaluation while it lasts. Confirmed, checked against the live OpenRouter model listing on 2026-08-23.
TL;DR - Last verified: 2026-08-23
- Ox Alpha (
stealth/ox-alpha) went live on OpenRouter on 2026-08-20 and costs $0 per million input and output tokens during preview. [Confirmed]- Specs: 1,048,576-token context, up to 131,072 output tokens, text + image + video input, mandatory reasoning mode. [Confirmed]
- The provider stays anonymous. Prompts are retained but stated not to be used for training. Do not send proprietary code or secrets.
- Independent forensics point to a Zhipu AI (Z.AI) GLM-family base. This is Reported, not confirmed by any lab.
- Free previews on OpenRouter have historically either become paid named models or disappeared, so treat access as borrowed time.
What is Ox Alpha, exactly?
Ox Alpha is a reasoning model listed on OpenRouter under the anonymous provider label "Stealth", model ID stealth/ox-alpha. OpenRouter's own description calls it "a reasoning model designed for coding, sustained agentic work, and production workloads". The listing went live on 2026-08-20 (creation timestamp 1787256295 on the OpenRouter model API).
A stealth model is a preview release where the actual lab behind the model stays unnamed. OpenRouter has run several of these before, and past stealth drops were later resolved to named models from major AI labs. The practical point: OpenRouter routes your traffic to whoever is serving the model, and you do not get to know who that is.
Confirmed specs from the live OpenRouter listing (checked 2026-08-23):
| Spec | Value |
|---|---|
| Model ID | stealth/ox-alpha |
| Price | $0/M input, $0/M output (preview window) |
| Context window | 1,048,576 tokens |
| Max output | 131,072 tokens |
| Input modalities | Text, image, video |
| Output modality | Text |
| Reasoning | Always on, default effort "max" |
Source: OpenRouter model listing, openrouter.ai/stealth/ox-alpha.
Is it actually good for coding?
Short answer: yes for its price (free), with an honest benchmark caveat.
An independent test posted publicly by a developer on 10 tasks from the DeepSWE benchmark showed Ox Alpha scoring ahead of some frontier models on that narrow slice, but a fuller run put it closer to roughly 63%, which groups it with strong open models rather than the very top of the closed frontier. This is a Reported, community-run benchmark, not an official eval. Treat any single benchmark post claiming it "beats frontier models" as incomplete.
What is more useful than a score is the shape of its output. In hands-on testing with full-stack app generation, Ox Alpha tends to pick a modern, sensible default stack and produce working, multi-page applications in a single long generation. Reasoning is always on and defaults to maximum effort, so generations are slow - tens of minutes for a large single-shot app build - but they cost nothing during the preview.
The practical read: Ox Alpha is a strong candidate for the "cheap, high-volume" tier in a routing setup - the model your agent harness burns tokens on - with a premium named model kept as the fallback for high-stakes merges and security-sensitive work.
Who built Ox Alpha?
Nobody official has said. The leading theory is that it runs on Zhipu AI / Z.AI infrastructure and is based on the GLM family, possibly a post-trained or multimodal successor to the GLM 5.3 line. This is Reported, not confirmed. The evidence that circulates for it:
- Tokenizer comparisons show an exact or near-exact match with known GLM tokenizers.
- The video-encoding token scheme matches the GLM vision model line.
- Malformed requests to at least one serving route returned error details consistent with Z.AI's serving stack.
Because the GLM family is open-weights, it is also possible a third party took GLM as a base and post-trained it - so "Ox Alpha" might not be Zhipu's own release at all. None of the companies involved have confirmed anything. Plan around the model's behavior, not its rumored identity.
What are the privacy terms and limits?
This is the part that decides whether you should send real work to it, and the words matter:
- OpenRouter's listing states prompts and completions are retained by the provider and stated not to be used for training. Retained means logged and kept by an anonymous operator. See the OpenRouter stealth model documentation for the governing terms.
- OpenCode, the open-source coding agent that also routes to Ox Alpha directly, advertised its route as "Zero Data Retention" with capacity for very large daily token volumes (Vendor claim, opencode.ai). That claim covers OpenCode's handling, which is a different party from the upstream provider OpenRouter refers to. Do not conflate them.
The safe rule with any free stealth preview:
- Do not send customer PII, credentials, API keys, or proprietary code you could not tolerate being logged.
- Do use it for open-source repos, sanitized codebases, spikes, benchmarks, and throwaway prototypes.
- Watch the model page and your usage dashboard. Past stealth previews on OpenRouter (Owl Alpha, Hunter Alpha and others) either graduated into named, paid models or vanished. "Free" here is a preview subsidy, and it can end without notice.
How do you use Ox Alpha today?
- Create a free API key at openrouter.ai.
- Point any OpenAI-compatible client at
https://openrouter.ai/api/v1with modelstealth/ox-alpha. This works in tools like Cline, Zed, and Claude Code's OpenRouter configuration. - Or install OpenCode and select Ox Alpha from its free models - no key juggling during the preview window.
Example API call:
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer <your-openrouter-api-key>" \
-H "Content-Type: application/json" \
-d '{
"model": "stealth/ox-alpha",
"messages": [{"role": "user", "content": "Refactor this Go connection pool for correctness"}]
}'
One operational note: reasoning is mandatory and defaults to "max" effort, so expect long response times on big prompts. If you use the API directly you can lower reasoning_effort to high or low for faster, cheaper-latency loops.
What this means for you
If you are a developer or running a small team, the best open-source LLM for coding answer this month is genuinely shifted by a free 1M-context model, even an anonymous one. Our broader coverage of the open-source coding model field and DeepSeek's latest vision update shows how fast this tier moves. Concretely:
-
Run your own evals now. Wire
stealth/ox-alphainto your agent harness and replay ten of your real tasks - lint fixes, migrations, test generation. A free week is exactly enough time to measure whether it earns a permanent slot. - Use it as the volume tier. Route cheap, high-frequency agent work to Ox Alpha and keep a named, enterprise-terms model for merges, security review, and anything with secrets in it.
- Do not build production dependencies on it. Preview models change price or disappear. Keep a fallback route configured from day one, and if you are newer to concepts like context windows and token pricing, our developer primer on LLM fundamentals will help you reason about the tradeoffs.
FAQ
Q: Is Ox Alpha free?
A: Yes, during its preview window Ox Alpha is listed at $0 per million input and output tokens on OpenRouter (checked 2026-08-23). Stealth previews have historically become paid or been removed after the preview, so check the model page before relying on the price.
Q: What is Ox Alpha's context window?
A: 1,048,576 tokens (1M) of context, with up to 131,072 output tokens per completion, according to OpenRouter's live model listing.
Q: Can Ox Alpha process images and video?
A: Yes. OpenRouter lists its input modalities as text, image, and video, with text output. That makes it one of the few free coding-capable models that accepts video input.
Q: Who made Ox Alpha?
A: It is a stealth model, so no lab is officially named. Community forensics (tokenizer matches, error fingerprints) point toward a Zhipu AI / GLM-family base, but that is reported analysis, not a confirmed identity.
Q: Is it safe to send my company's code to Ox Alpha?
A: No. OpenRouter states the provider retains prompts and completions (while stating they are not used for training), and the provider is anonymous. Stick to open-source, sanitized, or throwaway code until the lab is revealed and enterprise terms exist.
Q: Ox Alpha vs the paid frontier models - which should I use?
A: Use Ox Alpha for high-volume, non-sensitive agent work where cost dominates, and a named frontier model for high-stakes, sensitive, or correctness-critical steps. Community benchmarks place it near the strong open-model tier rather than at the very top, and those numbers are unofficial.
Sources
- Ox Alpha - OpenRouter model page (live listing, checked 2026-08-23; specs cross-checked against the OpenRouter models API)
- OpenRouter stealth model documentation (provider anonymity and retention terms)
- OpenCode - open-source AI coding agent (direct free routing to Ox Alpha; capacity claims are vendor claims)
Updates & Corrections
- 2026-08-23 - Initial publication. Specs verified against the live OpenRouter model API; GLM-base theory labelled as Reported; DeepSWE community figure labelled as unofficial.
Researched and drafted with AI agents; reviewed and fact-checked under human editorial oversight. See how we work.
Top comments (0)