Ox Alpha: Investigating the Free Anonymous Frontier Model

On August 20, 2026, a model with no company name, no press release, and no logo appeared on OpenRouter under the identifier stealth/ox-alpha. It offered a 1,048,576-token context window, accepted text, images, and video, claimed zero data retention, and was free for approximately one week. Within 48 hours, community researchers had produced significant forensic evidence pointing to a specific lab. This article covers what the model does, how well it performs, and what the investigation has found.

What Ox Alpha offers

OpenRouter describes it as "a reasoning model designed for coding, sustained agentic work, and production workloads" suited for "long-horizon software engineering, complex reasoning, and workflows that combine text with visual context." The free access runs through around August 27, 2026, and is available through both OpenRouter and the OpenCode Zen plan.

The stated capacity is striking: OpenCode claimed 100 trillion tokens per day, and Nous Research claimed 1 quadrillion tokens per day. Those numbers indicate a serious infrastructure commitment, not a small-scale experiment.

The zero data retention claim is on the per-model listing, but the anonymous provider's identity remains unconfirmed, and the governing terms are ambiguous. Don't use it with sensitive, proprietary, or personal data until the operator confirms their identity and data handling on the record.

Benchmark performance

An initial 10-task DeepSWE benchmark run by researcher Ben Davis produced an 80% score, placed ahead of Claude Fable at 65% and GPT-5.6-sol at 52%.

That 10-task subset is preliminary and not an audited leaderboard result. More complete testing by researcher Wenqi/Kevin produced a score around 63%, which is more representative.

A 63% on DeepSWE places Ox Alpha alongside DeepSeek-v4-pro (63%), GPT-5.5 high (64%), Grok-4.6 high (65%), and Gemini-3.7-flash high (65%). Day.dev testing placed it at 87.5% on Kingbench, behind GLM-5.3's 91.25% on the same benchmark.

The finance dashboard test

To investigate the model's identity, seven models were given the same prompt in an empty folder: build a personal finance dashboard with account tracking across savings, checking, investments, and debts; transaction history; spending analysis; financial goals; fund management; and seeded backend data. The prompt named no framework, language, or database. Each model had to make those choices independently.

Ox Alpha

Ox Alpha generated a complete multi-page dark-themed application with navigable sections for Overview, Accounts, Transactions, and Goals after about 45 minutes.

The technology stack in its package.json: Next.js, better-sqlite3, and Drizzle ORM. A real persistent database and a type-safe ORM, not an in-memory store.

The other models

DeepSeek v4 Pro produced a visually similar dark-themed UI. It also chose Next.js and better-sqlite3, but did not use Drizzle ORM.

Kimi K3 generated a plain React/Vite frontend with a separate Express backend and the more basic node:sqlite library.

Claude Fable Medium and GPT-5.6 Sol both produced light-themed UIs with hardcoded JavaScript objects instead of a database. Refreshing the page would lose all changes.

Gemini 3.7 Flash produced a dark-themed UI but also without persistent storage.

GLM-5.3 produced a light-themed UI with a package.json that was the decisive finding.

The smoking gun

GLM-5.3's technology stack: Next.js, better-sqlite3, and Drizzle ORM. Identical to Ox Alpha's.

Code forensics

Using an AI-assisted code attribution analysis, the two codebases were compared across the full set of models.

Structural similarity

Breaking code into overlapping token shingles and computing Jaccard overlap across all seven codebases:

The structural similarity score between Ox Alpha and GLM-5.3 is 0.224, the single highest off-diagonal value in the entire matrix.

Rare markers

Eight idiosyncratic coding choices were identified, things not requested in the prompt and not obvious defaults. Ox Alpha and GLM-5.3 shared all eight. No other pair shared more than two.

The shared markers included: an identical parseAmountToCents() function; use of Drizzle ORM; both models were the only ones to rewrite the README.md file with the same structure; both used asynchronous React Server Components; both used the exact same pnpx db:seed command via tsx.

Serving-layer forensics

Independent from the code analysis, researcher Chetaslua triggered a malformed request to OpenCode's Ox Alpha route on August 22, 2026. The returned Java stack trace named com.wd.paas.api.domain.v4.chat.ChatCompletionRequest, which is Zhipu's documented API class. The error code matched GLM-5.3's error code 1214. A tokenizer test found 30/30 identical matches between Ox Alpha and GLM-5.3.

Video encoder fingerprinting

Processing the same video files through Ox Alpha and GLM-5V-Turbo produced identical token counts in every row. Other models tested had different token usage patterns for the same inputs.

The current state of the investigation

The forensic evidence strongly points to Zhipu AI's GLM family. Zhipu has precedent for this: Pony Alpha, a stealth model released in February 2026, was later confirmed to be GLM-5. As of August 23, 2026, neither Zhipu nor OpenRouter has officially confirmed or denied anything. The leading theory is that Ox Alpha is a multimodal variant in the GLM-5.x line, possibly an unreleased GLM-5.5 or a post-trained version of GLM-5.3. The exact model version remains unsettled.

OpenRouter is in the news separately: Stripe recently acquired the platform, which adds some context to the ecosystem where Ox Alpha appeared.

One important caveat the forensics don't resolve: Zhipu AI operates under China's National Intelligence Law, which requires cooperation with state intelligence work. That obligation is the operative legal condition regardless of the model's capabilities or the zero data retention claim. Until the operator confirms their identity and data policies on the record, treat this as a capability evaluation tool only, not a production integration for sensitive workflows. The free window is a genuine opportunity to assess a likely frontier-class model at no cost. Keep proprietary code and personal data out of it.