Ox Alpha: What the Stealth Model Is and How to Set It Up (Before the Free Window Closes)

One-line job: Wire up a free 1M-context multimodal frontier model this week — and understand the data-terms contradiction before you paste anything into it. Audience: Developers who can run curl and want to evaluate Ox Alpha while it’s free. Not for: Anyone planning to send credentials, customer data, or NDA-covered material — this model’s data terms disqualify that use outright. Section 4 explains why. Last verified: 2026-08-21 Evidence weight: documentation-verified

On August 20, a frontier-class reasoning model appeared on OpenRouter under the provider name “stealth.” It has a million-token context window, takes text, images, and video, costs nothing during its preview week, and nobody has claimed it. The OpenCode launch announcement set the tone:

“Ox Alpha (stealth model) is free for the next week - 1M Context - Multi-modal - Zero Data Retention - Generous rate limits, near unlimited usage - We have capacity for 100T tokens per day, lets see what you can do” — OpenCode (@opencode), Aug 21, 2026

That capacity claim is the part that broke people’s brains. Theo of t3.gg:

“‘We have capacity for 100T tokens per day’ Okay who the fuck made this model and where did they get this much compute?” — Theo - t3.gg (@theo), Aug 21, 2026

The scale isn’t hypothetical. OpenRouter’s app table showed roughly 1.38 trillion prompt tokens consumed on day one, with Hermes Agent, Claude Code, omp, DeepSeek Harness, and ZCode as the top five consumers. This thing got used.

What it actually is

Per the OpenRouter model page:

“Ox Alpha is a reasoning model designed for coding, sustained agentic work, and production workloads.” — OpenRouter model page, 2026

Property	Value
Model ID	stealth/ox-alpha
Listed	Aug 20, 2026
Context / max output	1,048,576 / 131,072 tokens
Modalities	Text, images, video in; text out
Preview price	$0 / $0
Tool calling	Yes (tools + tool_choice)
Structured output	JSON via response_format, no schema enforcement
Throughput	P50 ~24 tok/s best-case; distribution avg ~39
Latency	~4.2s first token (best); E2E P50 ~16s
Tool call error rate	~2.37%
Cache hit rate	~80%+

Performance reality check

The informal numbers are promising: 87.5% on AICodeKing’s Kingbench (ahead of Opus 4.8 and Qwen 3.8 Max), and about 80% on Ben Davis’s ten-task Deep SWA subset — including one-shot solves where GLM-5.3, Grok 4.6, and GPT 5.6 all scored zero. But hold those numbers loosely, because both are single-tester private benchmarks. As Coursiv put it:

“A hand-picked sample of 10 can favor one model purely through task selection.” — Coursiv, Aug 2026

The official DeepSWE benchmark contains 113 tasks; a ten-task subset is a taste, not a verdict. Ox Alpha is absent from official leaderboards entirely — Startup Fortune’s framing (“The Numbers Are Real, The Leaderboard Is Not”) captures it. Reddit field reports are genuinely mixed: real bugs found that other tools missed, strong frontend and screenshot work, but also long stretches of overthinking without acting and struggles on complex backend tasks. Field-testing, not medal ceremony.

Setup — three routes

Route A: Playground (no code)

1. Go to openrouter.ai/stealth/ox-alpha.
2. Click Try this model / Playground. Create a free OpenRouter account if needed.
3. Submit text, an image, or a video prompt. Done — no key required for playground use.

Route B: API

OpenAI-compatible endpoint, so your existing client probably works with two changes:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "stealth/ox-alpha",
    "messages": [{"role": "user", "content": "Summarize this architecture diagram", "attachments": []}]
  }'

Details that matter:

• Streaming: "stream": true, standard SSE.
• Two extra endpoint shapes: it’s also exposed in OpenAI Responses format (POST /api/v1/responses) and Anthropic Messages format (POST /api/v1/messages — accepts text, images, PDFs, tools, extended thinking).
• The reasoning rule: when continuing a multi-turn conversation, send the complete reasoning_details array back to the model. Dropping it degrades the reasoning chain.
• Get your Bearer key from the OpenRouter dashboard.

Route C: OpenCode

Ox Alpha launched as a joint OpenCode promotion. Connect the OpenCode Zen provider, run /models, select Ox Alpha. Same free window, near-unlimited limits.

Operational expectations

Three things to know before you wire it into anything: median latency runs 3–4 seconds with ~24–39 tok/s throughput — noticeably slower than premium endpoints. The ~2.37% tool-call error rate means agent loops should wrap tool calls in retries. And the ~80% cache hit rate makes repeated-prefix agentic workloads unusually cheap. During the preview, everything is free anyway — which makes this week the cheapest serious eval target available.

A note on multi-turn reasoning

The reasoning_details rule deserves an example, because it’s the one thing most likely to bite people porting existing agents. When Ox Alpha answers, it returns a reasoning block plus a reasoning_details array alongside the visible completion. If your agent loop sends only the final text back as conversation history, the model loses its own reasoning thread and quality degrades on follow-ups — you’ll see it re-derive conclusions it had already reached, or abandon constraints it was holding. The fix is mechanical: capture reasoning_details from each response and replay the full array in subsequent requests. The same applies through the Anthropic Messages endpoint, where thinking blocks serve the equivalent role.

What “multimodal” means here

Video input is the differentiator — very few production endpoints take video at all. Practical uses that early testers found: feeding UI screen-recordings for frontend reproduction work, passing architecture-walkthrough videos for documentation drafting, and screenshot sequences for debugging sessions. Image handling covers diagrams, screenshots, and scanned documents. Output remains text-only in every case; there is no image generation.

Rate limits and capacity planning

“Near unlimited usage” during the free week still means the provider can end access without notice — EULA §2b again. If you’re running batch evaluations, checkpoint results continuously rather than assuming the endpoint survives your run. Uptime over the first three days measured 99.99% at best providers, but a preview week is not an SLA.

The fine print that matters most

Here is the section that should change how you use this model. Three sources directly contradict each other about what happens to your data:

First, the OpenRouter model page banner:

“This stealth model is developed and operated by a third-party model provider. Prompts and completions for this model are retained by the provider and are not used for training.” — OpenRouter model page banner, 2026

Second, the OpenRouter Stealth EULA — the legally operative document — says nearly the opposite:

“…your User Content may be collected by us and shared with the Stealth Provider… If you do not want your User Content to be provided to Stealth Providers for Stealth Model training, then you should refrain from accessing or using the Stealth Models.” — OpenRouter Stealth Program EULA §1, updated Jul 6, 2026

And it gets more explicit about the trade being made:

“In consideration for the provision of your User Content for Stealth Model training and improvement, access to the Stealth Models is provided to you free of charge.” — OpenRouter Stealth Program EULA §3

Third, OpenCode’s announcement claims “Zero Data Retention.”

These cannot all be true. The banner says retained-not-trained; the EULA licenses your content for training and names that license as the consideration for free access; OpenCode says nothing is retained at all. When a marketing claim and a signed agreement conflict, the agreement wins. The safe operating assumption: your inputs are stored by an unidentified party.

Which produces the actual rules for this week: never paste credentials, customer data, or anything under NDA. Treat every prompt as non-confidential. And note EULA §2b — stealth models may be removed “at any time, with or without notice” — so don’t build production dependencies during a free week, no matter how good the benchmarks look.

Verdict

Use it this week for exactly three things: evaluating whether its coding and reasoning profile fits your stack, long-context experiments on material you’d be fine publishing, and agentic workflows where you’ve wrapped tool calls in retries. Don’t use it for anything confidential, and don’t architect anything around it surviving. Community fingerprinting strongly suggests Zhipu AI behind it — our companion signal check walks that evidence — with a reveal expected when the free window closes around August 27. The pricing after that reveal is the real story; every prior stealth conversion landed aggressive-cheap.