Ox Alpha — often typed as 0x Alpha — appeared on OpenRouter on 20 August 2026 with no company name, no model card lineage and no press release. Within 48 hours it was one of the most talked-about releases of the month: a free, frontier-class reasoning model with a million-token context window, video input and a coding benchmark claim that put it ahead of models people actually pay for.
Nobody has officially said who built it. That combination — remarkable capability, zero cost, anonymous provider — is exactly the kind of thing a business should look at carefully before letting it anywhere near real work. This guide covers what Ox Alpha is, what the listing verifiably says, what the forensic evidence suggests about who is behind it, and how to trial a stealth large language model without handing an unknown operator your company data.
Table of contents
- What Is Ox Alpha?
- Ox Alpha Specifications: What the Listing Actually Says
- How Good Is Ox Alpha at Coding?
- Who Built Ox Alpha? The Forensic Evidence
- The Stealth Model Playbook
- What Free Really Costs: Data, Privacy and the Two Routes
- Should Your Business Touch Ox Alpha?
- How to Run a Sensible Ox Alpha Pilot
- What Happens After 27 August?
- Ox Alpha FAQ
- References
What Is Ox Alpha?
Ox Alpha is a stealth AI model: a frontier model published under a codename so that its creator can gather real-world usage data before attaching their brand to it. It appeared as stealth/ox-alpha on OpenRouter — the model marketplace recently in the news after Stripe’s acquisition of OpenRouter — and as a free model called Ox Alpha on OpenCode’s Zen endpoint, on 20 August 2026.
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 description reads like a positioning statement for a flagship launch — which is almost certainly what the free week is: a public dress rehearsal.
The scale of the giveaway is remarkable in its own right. OpenCode has said it has capacity for 100 trillion Ox Alpha tokens per day — around 1.16 billion tokens every second — and is offering near-unlimited usage for the whole preview window. Somebody is paying a very large compute bill to make this model free, which tells you how much a week of real-world developer traffic is worth to a frontier lab right now.
Why the anonymity?
Stealth releases let a lab benchmark against real developer traffic without the reputational risk of a named flop. If Ox Alpha performs well, the creator claims it and converts a week of goodwill into a launch. If it stumbles, the codename absorbs the damage and the lab quietly iterates. As we cover below, this playbook has run at least four times in 2026 alone, and every previous mystery model was eventually claimed by a major Chinese AI lab.
Ox Alpha Specifications: What the Listing Actually Says
Strip away the hype and the OpenRouter listing makes specific, checkable claims. The Ox Alpha specification sheet is genuinely unusual — very few models at any price combine a 1,048,576-token context window with video input and a 131,072-token output ceiling.
| Specification | Ox Alpha (as listed) |
|---|---|
| Context window | 1,048,576 tokens (1M) |
| Maximum output | 131,072 completion tokens |
| Input modalities | Text, images, video |
| Output | Text only |
| Tool calling | Yes, including tool_choice |
| Structured output | JSON via response_format, no schema enforcement |
| Pricing during preview | $0 prompt / $0 completion |
| Released | 20 August 2026 |
| Provider | “Stealth” — anonymous third party |
The scale of the context window
The output ceiling is one eighth of the input window — 131,072 tokens is 12.5% of 1,048,576. In practice that means Ox Alpha can read an entire mid-sized codebase or a stack of contracts in one request and still write a very long answer.
What a million tokens means in practice
A 1,048,576-token window is roughly 750,000 words of input — the equivalent of feeding the model a dozen novels, a full year of contracts, or a substantial monorepo in a single request. Workloads that previously required chunking, embeddings and retrieval pipelines can, at this scale, sometimes be replaced by simply pasting everything in. That does not make retrieval obsolete — long-context attention still degrades towards the middle of very large prompts — but it changes the economics of document-heavy tasks like due diligence review, log analysis and legacy code audits, and it is the capability that most separates Ox Alpha from the free models businesses are used to.
Serving performance
OpenRouter’s own telemetry for the model reports a P50 latency of 6.10 seconds, throughput of around 21 tokens per second and 99.99% uptime over its first three days. That is respectable for a heavily loaded free endpoint, but noticeably slower than paid frontier endpoints — worth knowing before you point an interactive product at it.
How Good Is Ox Alpha at Coding?
The claim that lit up social media is simple: 80% Pass@1 on a DeepSWE-style software engineering benchmark, ahead of the frontier models it was tested against. The figures being circulated put Ox Alpha at 80%, with Claude Opus 5 cited at 65% and GPT-5.6 cited at 52% on the same task set.
The caveats that matter
Those numbers deserve heavy qualification, and the more careful write-ups say so plainly. The 80% figure comes from a 10-task test run by an independent researcher, not an audited leaderboard. DeepSWE is a different benchmark from SWE-bench Verified, which is what most frontier labs quote — so the comparison is not apples to apples. Ten tasks is a sample size at which one lucky solve moves the score by ten points.
What testers agree on
Even with the asterisks, the qualitative consensus from the first 48 hours of testing is that Ox Alpha is genuinely strong at sustained, agentic coding work: it follows long multi-step instructions, uses tools competently and stays coherent deep into large contexts. Independent testing placed it among August 2026’s frontier models rather than the mid-tier free models developers are used to being given for nothing.
Who Built Ox Alpha? The Forensic Evidence
Officially, nobody knows. Unofficially, the AI community spent the weekend doing serving-layer forensics, and by 22 August 2026 the evidence pointed strongly in one direction: Zhipu AI, the Chinese lab behind the GLM model family, also known as Z.AI.
| Forensic signal | Finding | What it suggests |
|---|---|---|
| Java stack trace leak | Exposed class maps to Zhipu’s api/paas/v4 chat route | Served from Zhipu’s own infrastructure |
| Error dialect | Identical error code 1214 to Z.AI-hosted GLM models | Same serving stack, not just similar weights |
| Tokenizer fingerprint | 30/30 probe matches to GLM-5.3, plus a 75-token hidden wrapper | GLM-family tokenizer |
| Video encoding | Token consumption matches GLM-5V-Turbo across test sets | GLM-family vision pipeline |
| Audio behaviour | Rejects audio input, consistent with GLM-5V | Same modality boundaries |
How confident is the attribution?
One researcher put the operator-layer confidence at 0.98, and a widely shared analysis calls it 99% certain the model is a Zhipu GLM-5.x series build. The candidates that were tested and ruled out include Xiaomi’s MiMo v2.5, Qwen 3.8 Max and GLM-4.6V. But it is worth being precise about what the evidence shows: it identifies who is serving Ox Alpha, which is very strong but still circumstantial for who trained it.
The official silence
Zhipu has neither confirmed nor denied anything on the record, and OpenRouter and OpenCode have stayed silent too. That silence is part of the format — a prediction market on Manifold is literally trading on the answer. If the pattern of previous stealth drops holds, the reveal will come as a product launch within days of the free window closing.
The Stealth Model Playbook
Ox Alpha is not a novelty; it is the fifth act of a now-familiar 2026 pattern. Anonymous “alpha” models appear on routing platforms, run free for a week or two, and are then claimed as launches by major labs.
| Stealth codename | Appeared | Later revealed as |
|---|---|---|
| Pony Alpha | February 2026 | Zhipu AI — GLM-5 |
| Hunter Alpha | March 2026 | Xiaomi — MiMo-V2-Pro |
| Elephant Alpha | April 2026 | Ant Group — Ling-2.6-flash |
| Owl Alpha | Late April 2026 | Meituan — LongCat-2.0 |
| Ox Alpha | August 2026 | Unconfirmed — evidence points to Zhipu AI |
What the pattern tells you
Two things stand out. First, every named reveal so far has been a Chinese lab — stealth drops have become the standard Western go-to-market channel for Chinese frontier models, and Pony Alpha was Zhipu’s own GLM-5 launch vehicle six months ago. Second, the free window is the product: the lab is buying benchmark data, stress testing and word of mouth, and paying for it with compute. Understanding that trade is the key to using Ox Alpha sensibly.
What Free Really Costs: Data, Privacy and the Two Routes
There are two ways to use Ox Alpha for free, and they carry different data terms. This is the part most of the excited posts skip, and it is the part that matters most for a business.
| Factor | OpenCode Zen | OpenRouter |
|---|---|---|
| Model ID | x-preview-f-free | stealth/ox-alpha |
| Token price | $0 | $0 |
| Stated retention | Zero retention, no training on your data | Prompts and completions retained, not used for training |
| Governing terms | Provider product page | Stealth Model Terms, which grant broader collection rights |
| SLA | None | None |
Read the stricter document
Note the tension on the OpenRouter route: the model page says prompts are retained but not trained on, while the broader stealth agreement grants collection and training rights. When a product page and a general agreement differ, a cautious buyer follows the stricter boundary — assume anything you send may be kept and learned from. Remember what the free week is for: the lab wants your hardest real-world prompts as evaluation data, and authentic developer sessions are exactly the raw material that modern post-training and reinforcement learning pipelines are hungry for.
What never goes in
The practical rule writes itself: no secrets, no credentials, no regulated data, no customer repositories, no personal, health or financial records. If a workload cannot leave your estate on those terms, the answer is a private AI deployment where the model comes to your data, not an anonymous free endpoint — however good it is.
Should Your Business Touch Ox Alpha?
For most UK SMEs the honest answer is: yes for evaluation, no for production — at least this week. The upside of a free frontier-class coding model is real, and so are the reasons it cannot be a production dependency yet.
Where Ox Alpha makes sense right now
It is a superb zero-cost benchmark. If you are choosing a coding model, evaluating agents, or costing an AI project, a week of free access to a frontier-class model lets you calibrate what “good” looks like without burning budget. Throwaway prototypes, open-source contributions, synthetic-data experiments and internal tooling on non-sensitive code are all reasonable uses. It is also a useful yardstick if you are weighing an off-the-shelf copilot against a custom AI assistant.
Where it does not
Anything customer-facing, anything with an uptime requirement, and anything touching data you would not post publicly. The provider is anonymous, there is no support channel, no SLA, no way to pin the weights, and the endpoint can change or vanish without notice. Free previews of stealth models are routinely slowed, constrained or withdrawn. Building on Ox Alpha today is building on sand — knowingly, for the right experiments, that is fine; unknowingly, it is an outage waiting to happen.
How to Run a Sensible Ox Alpha Pilot
If you want the free week to produce a decision rather than anecdotes, treat the Ox Alpha trial like a compressed vendor evaluation.
Set it up defensively
Create a separate API key with the lowest practical budget cap, and archive a copy of the terms as they stand today. Point it at a sanitised copy of a repository with secrets stripped — never a live customer codebase. Watch for hidden costs: model tokens are free, but PDF parsing, web search and external tool calls can bill separately.
Test what you actually do
Pick 15 to 25 tasks that represent your real workload — bug fixes, refactors, document extraction, long-context review — and run each task class three times to expose variance. Add a few adversarial prompts to see how it behaves at the edges. Score it like a supplier: accepted changes, defect rate, latency, and any non-model charges.
Compare against a paid baseline
Run the same task set through whichever model you currently pay for. The interesting question is not “is Ox Alpha good?” but “is it enough better to justify waiting for its real identity, price and terms?” Keep notes — when the model is claimed and priced, your evaluation converts directly into a purchasing decision. Our AI models and tools hub tracks the models worth benchmarking against.
What Happens After 27 August?
The free window was announced as roughly one week, ending around 27 August 2026, and post-preview pricing is listed as TBD. The likeliest outcome, given the four precedents, is a named launch: Ox Alpha resurfaces as a numbered GLM release with real pricing, real documentation and an accountable operator — at which point the anonymity discount disappears and normal vendor due diligence applies.
The strategic takeaway is bigger than one model. Frontier-class capability is now being given away in week-long windows to buy mindshare, and the gap between “best model money can buy” and “best model you can get for nothing this week” has never been narrower. Businesses that keep a lightweight evaluation harness ready — task list, scorecard, sanitised test repo — can turn each of these windows into free market intelligence. The ones that paste production data into every shiny new endpoint will eventually pay for the free week many times over.
Ox Alpha FAQ
Is Ox Alpha the same thing as 0x Alpha?
Yes. The official listing spells it Ox Alpha — the letter O, not a zero — but because the name is usually rendered in lowercase as stealth/ox-alpha, a large share of people searching for it type 0x Alpha, borrowing the hexadecimal prefix from programming. Either spelling gets you to the same mystery model.
Is Ox Alpha really free?
Model tokens are genuinely priced at zero on both routes during the preview, which runs until around 27 August 2026. The costs to watch are non-model charges — PDF parsing, web search and external tool calls can bill separately — and the non-monetary price: on the OpenRouter route your prompts are retained, and the stealth terms grant broad data rights. Free means the lab is being paid in your usage data.
Is Ox Alpha better than Claude or GPT?
Unproven. The circulated numbers — 80% for Ox Alpha against 65% and 52% for the cited rivals — come from a ten-task community test on a benchmark the big labs do not quote. The fair summary is that it is credibly frontier-class at coding and long-context work, and that audited head-to-head numbers do not yet exist. Run your own tasks through it and trust your own scorecard.
Who owns the code Ox Alpha writes for me?
Nobody can say with certainty, and that is precisely the problem: the provider is anonymous and the stealth terms are thinner than a normal commercial licence. For open-source experiments and internal throwaway tooling the risk is small. For anything you intend to ship commercially, wait until the model has a name, a licence and an accountable operator behind it.
Will Ox Alpha disappear?
The endpoint almost certainly will — stealth models are routinely withdrawn when the preview ends. The model itself will most likely reappear under its real name, probably as a GLM-series release if the forensic evidence is right, with real pricing attached. Plan any pilot so it produces a decision you can act on after the reveal, not a workflow that dies with the free window.
References
Ox Alpha — OpenRouter model listing
Ox Alpha: What We Know — Serving Layer Evidence — explainx.ai
Ox Alpha on OpenRouter: Free 1M Stealth Model — explainx.ai
OX Alpha: The Anonymous Frontier Model — Comprehensive Technical Analysis — Local AI Zone
Ox Alpha Free AI Model: Setup, Privacy and Cost — Wavect
A Mystery Model Called Ox Alpha Just Topped Coding Benchmarks for Free — Startup Fortune
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.