Ox Alpha Stealth Model: Who Made It and How to Use It

Three days after it appeared, Ox Alpha is the most-discussed model on OpenRouter and the least-documented. Nearly every article about it repeats the same headline number — "80% on DeepSWE, beats GPT-5.6" — as though it were an audited benchmark. It isn't. This page separates what the listing actually says, what one developer actually measured, and what is still guesswork. Volatile facts are date-stamped; everything below was checked against OpenRouter's live model page and public API on 23 August 2026.

What is Ox Alpha?

Ox Alpha is a reasoning model published on OpenRouter under the provider name "Stealth." OpenRouter's listing states the model is developed and operated by a third party that "has chosen to remain anonymous," and that OpenRouter is neither its developer nor its owner. It is described as built for "coding, sustained agentic work, and production workloads," and specifically for "long-horizon software engineering." Here is the listing, straight from the model page and the public /api/v1/models endpoint:

Two rows matter more than the headline 1M context. Reasoning is mandatory and defaults to max effort, which is why latency looks high. And there is no structured-output enforcement — a real constraint for agents that expect schema-valid JSON. The performance stats also drift: throughput sat near 25 tokens/second at the time of writing, and other observers reported roughly double a day earlier. Treat any single throughput figure for a stealth endpoint as a snapshot, not a spec.

Who made Ox Alpha?

Nobody has said. As of 23 August 2026 no lab has claimed Ox Alpha, and Zhipu AI — the company most people are pointing at — has not commented.

The leading theory is that Ox Alpha is an unreleased multimodal member of Zhipu/Z.ai's GLM-5.x family. It rests on fingerprinting, not on any lab statement. The evidence, mostly assembled by independent developer Ben Davis (@davis7) and echoed by a community probe tool called modelprint:

• Tokenizer. Across 25 test prompts, Ox Alpha's token counts matched GLM-5.3 exactly, with a constant +75-token offset attributed to a hidden system wrapper.
• Video encoder. Ox Alpha's video token consumption matched GLM-5V-Turbo across controlled samples — same frame-sampling behaviour, same duration scaling (~147 tokens per second of video), same resolution mechanics. Other candidate models diverged.
• Audio rejection. It refuses audio input the way GLM-5V does. Xiaomi's MiMo v2.5, one of the alternative candidates, accepts audio — which weakens that theory.
• Output style. Roughly 1.3 emoji per 1,000 characters, in line with GLM and Qwen outputs and far above the near-zero rate of Claude, GPT and Grok models.
• Error surface. Observers reported GLM-characteristic error codes leaking through on failure.

Davis put his own confidence at "99%." That is his estimate of his own inference, not a confirmation.

One more data point: the API metadata is a near-exact match

Here is a check you can run yourself, and one we haven't seen elsewhere. Pull both models from OpenRouter's public models endpoint and compare the configuration blocks:

curl -s https://openrouter.ai/api/v1/models \
  | python3 -c "import sys,json;[print(json.dumps({k:m[k] for k in ('id','supported_parameters','default_parameters','reasoning')},indent=2)) for m in json.load(sys.stdin)['data'] if m['id'] in ('stealth/ox-alpha','z-ai/glm-5.3','z-ai/glm-5.2')]"

As of 23 August 2026, stealth/ox-alpha and z-ai/glm-5.3 return an identical parameter surface: the same ten supported parameters, the same defaults (temperature: 1, top_p: 0.95), the same context length and max completion tokens, and the same unusual reasoning configuration — mandatory, always enabled, efforts max/high/low, defaulting to max. Both also omit structured_outputs.

That last detail is the interesting one. z-ai/glm-5.2 does support structured_outputs, plus a dozen parameters (seed, logprobs, stop, penalties) that neither GLM-5.3 nor Ox Alpha expose, and its reasoning block differs again (optional, efforts high/xhigh). So the match isn't a generic "Z.ai shape" — it is specific to the 5.3 generation. The missing JSON-schema enforcement people are calling an Ox Alpha limitation is actually a GLM-5.3-generation trait.

What this does not prove. Parameter metadata is set by whoever configures the endpoint, and a lab reusing its own serving stack would naturally produce a matching block. It is consistent with the GLM theory; it is not proof, and it cannot rule out another lab running a similar configuration.

The complication for the theory: GLM-5.3 — listed two days earlier, on 18 August 2026, at $1.40/$4.40 per million tokens — is text-only. Ox Alpha takes images and video. If the attribution holds, Ox Alpha is not GLM-5.3 itself but an unreleased multimodal sibling, which is why community shorthand has landed on names like "GLM-5.3V" or "GLM-5.5." Those names are speculation. Our GLM-5.3 launch guide covers what has actually shipped from that family.

How good is Ox Alpha really?

The number everyone quotes is 80% on DeepSWE. Here is what that number actually is. Ben Davis ran Ox Alpha through a 10-task subset of DeepSWE, a long-horizon software-engineering benchmark, and it passed 8 of 10. That is the 80%. It is not the full benchmark — DeepSWE's public set is far larger — and it is not an audited leaderboard entry. Davis flagged the limitation himself, noting sample sizes varied across the models he compared and that Ox Alpha's was small.

The comparison figures deserve the same scrutiny. Check the arithmetic: on a 10-item test every score must be a multiple of 10, so the reported 65% for Claude Fable 5 and 62% for GLM-5.3 and Grok 4.6 are impossible on 10 tasks. Those baselines come from different, larger runs. The models below were not all measured on the same sample.

Every cell in that table is community-sourced. There are no vendor-published benchmarks for Ox Alpha, because there is no vendor. Artificial Analysis does not list the model at all, it does not appear on public DeepSWE leaderboards, and OpenRouter's model page shows no benchmark block.

What the result does tell you: on at least one hard agentic-coding sample, Ox Alpha completed tasks that several frontier models failed, with a low retry rate on a long tool-calling run. That is a genuine signal. What it does not tell you is where the model lands on SWE-bench Verified, on reasoning or math suites, on multilingual work, on long-context recall at 1M tokens, or on refusal and safety behaviour. Those are simply unmeasured. And quoting Ox Alpha's DeepSWE 80% next to another model's SWE-bench Verified 96% compares two different exams — for a like-for-like frontier reference point, our Kimi K3 benchmark comparison uses consistent suites across models.

How do you use Ox Alpha right now?

The model is exposed through OpenRouter's OpenAI-compatible API. Get a key from your OpenRouter dashboard and point any OpenAI SDK at https://openrouter.ai/api/v1.

curl:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "stealth/ox-alpha",
    "messages": [
      {"role": "user", "content": "Refactor this module and explain the tradeoffs."}
    ],
    "reasoning": {"effort": "max"},
    "max_tokens": 8192
  }'

Python, via the OpenAI SDK:

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)

resp = client.chat.completions.create(
    model="stealth/ox-alpha",
    messages=[{"role": "user", "content": "Summarise this repo's failure modes."}],
    extra_body={"reasoning": {"effort": "max"}},
    max_tokens=8192,
)
print(resp.choices[0].message.content)

Three notes that will save you a debugging hour:

• Reasoning cannot be disabled. Drop effort to low or high to cut latency, but there is no non-reasoning mode.
• response_format works for "give me JSON," but there is no schema validation — parse defensively and retry on malformed output.
• OpenRouter's documented free-variant rate limits (50 requests/day under 10 credits purchased, 1,000/day above, 20/minute) apply to IDs ending in :free. stealth/ox-alpha is a zero-priced model, not a :free variant, so those caps don't govern it. Preview capacity is set by the anonymous operator and can change without notice.

Beyond the direct API, the model was distributed to partner surfaces during the preview: OpenCode (via its OpenCode Zen free plan), Cline and Venice all opened free access. OpenCode stated it had capacity for 100 trillion tokens per day — the clearest signal of how much scale the anonymous operator is buying for this test.

What are the catches?

Four, and the first is the one to take seriously.

1. Your prompts go to an anonymous operator. OpenRouter's listing is explicit that prompts and completions are retained by the provider, and states they are not used for training. But you cannot audit a company you cannot name: no published retention period, no jurisdiction, no data-processing agreement, no reviewable security posture, no counterparty if something goes wrong. OpenRouter has been direct about this pattern before — its 2025 Quasar Alpha announcement said outright that all prompts and completions would be logged so the lab could understand how the model was being used.

The practical rule: do not send proprietary source code, customer data, credentials or anything under an NDA to Ox Alpha. Use it on open-source repositories, scratch projects, synthetic data and public problems. Note too that OpenCode advertises "zero data retention" for its own client layer — a claim about OpenCode's infrastructure, not the upstream model provider, and it does not override retention at the model layer. Check the whole path, not one hop of it.

2. Throughput is slow for a frontier model. At roughly 25 tokens/second with mandatory max-effort reasoning and P50 latency near 5 seconds, Ox Alpha is a poor fit for anything interactive. Fine for batch agent runs you come back to; painful in an IDE autocomplete loop.

3. No JSON-schema enforcement. If your agent depends on guaranteed output shape, budget for a validation-and-retry layer.

4. It can vanish. A stealth listing has no SLA and no deprecation policy. Quasar Alpha and Optimus Alpha disappeared when OpenAI revealed them as GPT-4.1 prereleases; Sonoma Sky and Dusk Alpha disappeared when xAI revealed them as Grok 4 Fast. Never make it the only model in a production path.

What happens when the free window ends?

Worth being precise, because most coverage is not. OpenRouter's model page carries no stated end date for the free pricing. The "one week" figure everyone quotes comes from partner announcements — OpenCode said "free for the next week" on 20 August — and press coverage extrapolating to roughly 27 August. As of 23 August 2026 the listing still shows $0 in and $0 out.

Three plausible endings, all with precedent: the model is unmasked and folded into its lab's paid catalogue under a real name; pricing is switched on at the same endpoint; or the listing is pulled with no announcement.

If the GLM theory is right, the natural successor is Z.ai's own catalogue. GLM-5.3 sits at $1.40 in / $4.40 out per million tokens with the same 1M context and 131K output ceiling — so if you are prototyping on Ox Alpha, the cheapest way to de-risk is to build against that price point and confirm your workload is still viable there. Our GLM-5.2 guide covers the open-weights side of that family, and the open-source LLM landscape pillar maps the alternatives.

Why do labs launch stealth models at all?

Anonymous previews on OpenRouter are now an established launch pattern, not a novelty. The two clearest precedents:

• Quasar Alpha and Optimus Alpha (April 2025) — both turned out to be early versions of OpenAI's GPT-4.1. OpenRouter described Quasar Alpha at the time as "a prerelease of an upcoming long-context foundation model (1M tokens) from one of the model labs."
• Sonoma Sky Alpha and Sonoma Dusk Alpha (September 2025) — xAI's Grok 4 Fast, in reasoning and non-reasoning variants.

What the lab gets tells you what you are actually participating in:

• Unbranded evaluation. Developers judge output quality without knowing whose logo is on it — no halo effect, no backlash, no prior shaping the reaction.
• Load testing at real scale. Free tokens buy production-shaped traffic — long agent loops, odd tool schemas, adversarial prompts — that no internal eval reproduces. OpenCode's 100-trillion-tokens-per-day capacity note is the giveaway that this is a scale test.
• Competitive telemetry. Real developer prompts across real codebases: the exact distribution a coding model needs, and the hardest thing to buy.
• A launch narrative. A week of "who made this?" is cheaper and more credible than an ad campaign, and the reveal lands with an audience already convinced.

None of that is sinister, but it is a trade: you get free frontier-class inference, the lab gets your usage patterns. Priced honestly, that is reasonable for non-sensitive work.

Should you use Ox Alpha right now?

Use it if you are running batch agentic coding tasks on public or non-sensitive code, you want a free look at a possible frontier model, and you have a fallback wired in. Skip it if the work involves client code, customer data or anything under contract; if you need low latency or schema-guaranteed JSON; or if it would become a product dependency on an endpoint with no SLA that may be gone next week.

The most useful thing you can do in the remaining window is run your own evaluation on your own tasks. A 10-task community result is a hint. Twenty tasks from your actual backlog will tell you more about whether this model is worth paying for later than every benchmark table currently in circulation.

FAQ

What is Ox Alpha?

Ox Alpha is an anonymous "stealth" AI model listed on OpenRouter as stealth/ox-alpha since 20 August 2026. It is a reasoning model aimed at coding and long-horizon agentic work, with a 1,048,576-token context window, 131,072-token maximum output, and text, image and video input. OpenRouter routes requests to it but is not its developer or owner.

Who made Ox Alpha?

Officially unknown. As of 23 August 2026 no company has claimed it, and OpenRouter states the operator has chosen to remain anonymous during the preview. The leading community theory attributes it to Zhipu AI (Z.ai), the lab behind GLM, based on tokenizer, video-encoder and behavioural fingerprinting. Zhipu has not commented.

Is Ox Alpha free?

Yes as of 23 August 2026 — OpenRouter lists it at $0 input and $0 output. The nuance most coverage misses: OpenRouter publishes no end date for that pricing. The widely quoted "one week from 20 August" comes from partner announcements such as OpenCode's, not from OpenRouter. Check the live listing before relying on it.

Is Ox Alpha actually GLM-5?

Unconfirmed. The evidence is strong but circumstantial: an exact tokenizer match with GLM-5.3 across 25 prompts, video-encoder behaviour matching GLM-5V-Turbo, GLM-style audio rejection, and an identical API parameter and reasoning configuration to GLM-5.3 on OpenRouter. One inconsistency: GLM-5.3 is text-only while Ox Alpha is multimodal, so it would have to be an unreleased sibling.

Is Ox Alpha safe to use with private code?

No. OpenRouter's listing states prompts and completions are retained by the anonymous provider. Even with the "not used for training" assurance, you cannot audit an unnamed company, and there is no published retention period, jurisdiction or counterparty for a data agreement. Restrict it to open-source repositories, scratch projects and synthetic data.

How do I access Ox Alpha?

Through OpenRouter's OpenAI-compatible API using model ID stealth/ox-alpha at https://openrouter.ai/api/v1/chat/completions with a Bearer token from your OpenRouter dashboard. During the preview it was also distributed to partner surfaces including OpenCode (via the OpenCode Zen free plan), Cline and Venice, each with its own limits and data terms.

How good is Ox Alpha at coding?

Promising but under-measured. The 80% DeepSWE figure comes from one developer, Ben Davis, passing 8 of 10 tasks on a small subset — not the full benchmark and not an audited leaderboard. Artificial Analysis does not list the model at all. On that sample it completed tasks several frontier models failed, which is a real signal, but its performance on standard suites is unknown.

What happens when the Ox Alpha free window closes?

One of three things, all with precedent: the model is unmasked and moved into its lab's paid catalogue under a real name, pricing is switched on at the same endpoint, or the listing disappears without notice. Quasar Alpha and Optimus Alpha became GPT-4.1; Sonoma Sky and Dusk Alpha became Grok 4 Fast. Keep a paid fallback configured.