Ox Alpha Model 2026: Is the Free Unlimited AI Real?
Ox Alpha is free, has a 1M context window, and claims 100 trillion tokens a day of capacity. Only one of the two routes is actually unlimited, and only one keeps your prompts private.
16 min read
A stealth model called Ox Alpha showed up on August 20, 2026, with a 1 million token context window, image and video input, and a price tag of exactly zero. Within hours people were reporting billions of tokens burned in a single afternoon.
Ox Alpha is a real frontier-class model that is genuinely free for about one week, but "unlimited" only describes one of the two ways you can reach it. On OpenCode Zen the limits are effectively wide open. On OpenRouter you are capped at 20 requests per minute and 1,000 requests per day. Same model, same weights, completely different experience. The route you pick matters more than the model does.
I did not run a benchmark on it. I read both providers' own documentation side by side and did the arithmetic, and the two pages do not say the same thing. That gap is the whole story.
Key takeaways
- Ox Alpha is a stealth model on OpenRouter (
stealth/ox-alpha) and OpenCode Zen (x-preview-f-free). 1,048,576 token context, 131,072 max output, text plus image plus video in. - Free for roughly one week from August 20, 2026. OpenCode claims capacity for 100 trillion tokens per day.
- OpenRouter's free tier caps you at 20 requests per minute and 1,000 requests per day if you have spent at least $10 lifetime, or 50 per day if you have not.
- Data retention is not the same on both routes. OpenCode's docs say zero retention and no training. OpenRouter's page says prompts and completions are retained by the provider.
- Nobody knows which lab built it. The popular guess is a big American lab. The base rate says otherwise.
What is Ox Alpha and why is everyone burning billions of tokens on it?
Ox Alpha is what the industry calls a stealth model. A lab ships a model under a codename through a router, watches how real developers use it, collects the failure cases, then either reveals itself or quietly kills the endpoint.
This is not new. In September 2025, two models called Sonoma Dusk Alpha and Sonoma Sky Alpha appeared on OpenRouter with 2 million token context windows and free access. Everyone speculated for weeks. They turned out to be early builds of Grok 4 Fast from xAI.
So the pattern is established. Free frontier model with no name, one to three weeks of open season, then a reveal.
What makes Ox Alpha different is the scale of the offer. OpenCode publicly said they have capacity for 100 trillion tokens per day on it. That number is doing a lot of work in people's heads, and I want to take it apart, because it is not the flex it sounds like.
Is Ox Alpha actually unlimited?
No. It depends entirely on which door you walk through.
Here is the same model on both routes, pulled from each provider's own pages:
| OpenCode Zen | OpenRouter | |
|---|---|---|
| Model ID | x-preview-f-free | stealth/ox-alpha |
| Requests per day | No published cap | 1,000 (or 50 under $10 lifetime credits) |
| Requests per minute | No published cap | 20 |
| Throughput | ~50 to 60 tokens/sec reported | 34 tokens/sec, 3.74s P50 latency |
| Data retention | Zero retention, no training | Retained by provider, not used for training |
| Auth | Sign in, add billing, get API key | API key |
Read that throughput row again. OpenRouter publishes 34 tokens per second and a 3.74 second P50 latency on their own model page. Developers running it through OpenCode are reporting 50 to 60 tokens per second. That is roughly a 1.5x speed difference on identical weights, purely from routing.
The rate limit row is worse. Twenty requests per minute means you can spend your entire daily allowance in fifty minutes and then sit on your hands until tomorrow.
The model is not rate limited. You are.
[Image: Side by side comparison of the OpenCode Zen model page and the OpenRouter Ox Alpha page, with the rate limit and data retention lines highlighted]
Why does OpenRouter cut you off at 1,000 requests?
Because OpenRouter's free tier limits have nothing to do with backend capacity. They are a spam gate.
Their documentation is blunt about it. Free models get 20 requests per minute across the board. Daily allowance is 50 requests if you have purchased less than $10 in credits over the lifetime of your account, and 1,000 requests if you have crossed $10. They also state that creating extra accounts or API keys does not help, because capacity is governed globally.
So the gate is on your wallet history, not on the model's ability to serve you.
Now do the math on how absurd that is next to the 100 trillion number. If you hit the wall at around 10 million tokens, you consumed 0.00001 percent of the claimed daily capacity before being told to come back tomorrow.
And a modern agentic coding session is not one request. A single task where the model reads files, runs a grep, edits three files and reruns a test is easily 30 to 60 round trips. At that rate 1,000 requests buys you somewhere between 17 and 33 real sessions in a day.
That is not unlimited. That is a generous demo.
What would those free tokens actually have cost you?
This is where it gets fun, and this is the number nobody is computing.
One developer publicly reported burning roughly 2.72 billion tokens in about three hours. The split was 2.66 billion cache tokens, 44.4 million input tokens, and 17 million output tokens.
Look at that split. 97.7 percent of it was cache reads.
Now price the same workload on paid models, using OpenCode Zen's published rates:
| Model | Cached read | Input | Output | Total for 2.72B tokens |
|---|---|---|---|---|
| DeepSeek V4 Pro (off-peak) | $0.022/M | $0.66/M | $1.98/M | ~$121 |
| GLM 5.2 | $0.26/M | $1.40/M | $4.40/M | ~$829 |
| Claude Opus 5 | $0.50/M | $5.00/M | $25.00/M | ~$1,977 |
Three hours of one person's evaluation run. Between one hundred and two thousand dollars depending on which model you would have used instead.
But here is the part that reframes the entire "100 trillion tokens per day" boast. Cache reads are the cheapest thing a provider serves. On DeepSeek V4 Pro the cached read price is $0.022 per million against $0.66 for fresh input. That is a 30x gap, and it exists because a cache hit skips almost all of the expensive prefill compute.
If that same 2.66 billion had been fresh input instead of cache reads, the GLM 5.2 bill goes from about $829 to about $3,724. On Opus 5 it goes from $1,977 to roughly $13,300.
So when a provider says they can serve 100 trillion tokens a day, they are quoting a number that is dominated by the cheapest possible token. Agentic harnesses re-send the same system prompt and the same file context on every single turn. Almost all of that is cache. The headline capacity number and the actual GPU bill are not close to each other.
That does not make the offer fake. It makes it affordable, which is exactly why it is happening.
Does the zero data retention promise actually hold?
Partially, and the fine print is not identical on both routes. This is the single most important thing in this post if you are doing anything sensitive.
OpenCode's Zen documentation lists a zero retention policy across their providers, then lists exceptions. Several of their free models are explicitly carved out with language like "during its free period, collected data may be used to improve the model." Big Pickle, MiMo-V2.5 Free, Hy3 Free and the NVIDIA Nemotron endpoints all carry that warning.
Ox Alpha Free does not. Its line says the provider follows a zero retention policy and does not use your data for model training. Good.
Now open OpenRouter's page for the same model. It says prompts and completions are retained by the provider and are not used for training, with everything else governed by their Stealth Model Terms.
Not used for training is a different promise from not stored. One says your data will not shape the next model. The other says your data is not sitting on someone's disk. Only OpenCode is making the second promise.
And remember who the provider is here: nobody knows. It is an anonymous third party.
People are excited that this model has weak safety guardrails and will happily reason about exploitation. That is genuinely useful for security work. But pair the two facts honestly:
An anonymous provider, no guardrails, and a route where your prompts are retained. If you are pasting real infrastructure details into a security workflow, the route you chose just became a security decision.
I run artofcode on Convex and Next.js, and none of the 22 tools on the site call an LLM in the hot path, so I have no production traffic riding on this. But I would not put a client's network topology through an unnamed endpoint on the retained route, free or not.
Who is actually behind Ox Alpha?
Unknown as of August 21, 2026. The popular theory in developer circles is that only a large American lab could afford this, so it must be Google or similar.
I think that reasoning is the weakest part of the hype, for three reasons.
One: the giveaway precedent is Chinese, not American. DeepSeek V4 Flash processed 8 trillion tokens on OpenCode on August 1, 2026, and 5 of those 8 trillion were free trial usage. A Chinese lab already ate five trillion tokens of free traffic in a single day, three weeks ago. The premise that Chinese labs would not do this is factually wrong, because one of them just did.
Two: the anonymous-release base rate points the other way. Recent cloaked launches on these routers have come from Zhipu AI, Xiaomi, Ant Group and Meituan. The one big Western exception was xAI with the Sonoma pair. If you are betting on base rates, you bet Chinese lab.
Three: the routing tell everyone will reach for proves nothing. On OpenCode Zen, every Google model routes through a Google-specific endpoint with @ai-sdk/google. Ox Alpha Free routes through the generic /v1/chat/completions path with @ai-sdk/openai-compatible, the same path as DeepSeek, GLM, Kimi and MiniMax. That looks like evidence against the Gemini theory. It is not, and I want to be honest about that: a lab hiding its identity would route through the generic path on purpose. The absence of a fingerprint is what anonymity looks like. It tells you nothing.
Some analysts have published tokenizer fingerprints pointing at an unreleased GLM variant. Treat that as a rumor until someone shows the raw token counts. There is currently no independent benchmark sheet for this model at all. No intelligence index, no coding score, no agentic score.
Do not make architecture decisions based on a guessed identity. Your own tasks are the only evaluation that counts this week.
Why do labs hand out frontier models for free?
Three things they cannot buy any other way.
Load testing at real scale. You cannot simulate ten thousand developers running agentic loops against your inference stack. You have to actually get hit. A free week is the cheapest stress test in existence.
Failure cases from real work. Synthetic evaluations plateau. Real developers throw genuinely weird stuff at a model, and the traces from a week of that are worth more than another month of internal benchmarking.
Harness position. Once a model is the default in your opencode.json, switching costs go up. Being free during the trial is how you get into the config file. Being good enough is how you stay there after the price appears.
Notice that none of those three require the model to stay free. They all require it to be free right now.
How should you actually use Ox Alpha this week?
Use it, but use it with an expiry date written into your config so it fails loudly instead of silently.
Set the free model as your default and keep a cheap paid model for the small stuff:
{
"$schema": "https://opencode.ai/config.json",
"model": "opencode/x-preview-f-free",
"small_model": "opencode/deepseek-v4-flash"
}
If you are calling the API yourself, hard code the window so your code tells you when it closes instead of quietly erroring at 3am:
// model-router.js
const FREE_WINDOW_ENDS = new Date("2026-08-27T00:00:00Z");
const MODELS = {
free: "opencode/x-preview-f-free",
batch: "opencode/deepseek-v4-flash",
interactive: "opencode/claude-sonnet-5",
};
export function pickModel({ interactive, now = new Date() }) {
// Never put an unnamed provider in front of a user typing live.
if (interactive) return MODELS.interactive;
if (now >= FREE_WINDOW_ENDS) {
console.warn("[router] Ox Alpha free window closed, falling back to paid batch model");
return MODELS.batch;
}
return MODELS.free;
}
Two rules I would follow with any stealth model:
- Batch work only. Refactors, test generation, doc sweeps, migration passes. Things where a bad answer costs you a re-run and nothing else.
- Nothing confidential on the retained route. If you must use OpenRouter for it, assume the prompt is stored.
If you are queueing bulk jobs to run inside the window, a scheduled sweep beats sitting there babysitting it. I built a cron expression generator for exactly this kind of thing, because I could never remember the field order either.
And if you are pushing near the 1 million context ceiling, check the size of your prompt before you send it. A rough working ratio is about 1.3 tokens per word for English, so a 1 million token window is somewhere around 750,000 words of room. Most people who think they are hitting a context limit are actually hitting a retrieval problem.
[Image: Terminal screenshot showing the opencode.json config with x-preview-f-free set as the default model]
What breaks when the free window closes?
Around August 27, 2026, one of three things happens.
It gets revealed and priced. Most likely. If the GLM guesses are right, GLM 5.2 on Zen sits at $1.40 input and $4.40 output per million. That is your rough expectation. Cheap, not free.
It disappears entirely. Also common. The endpoint 404s and your agent loop dies at whatever step it was on. This is why the expiry date belongs in your config.
It stays free with real limits. Least likely, and if it happens the limits will look a lot more like OpenRouter's 1,000 requests than like this week.
Plan for the second one. Hope for the first.
What to remember
- Ox Alpha is real and good enough to be a daily driver for batch coding work this week.
- "Unlimited" applies to OpenCode Zen only. OpenRouter caps you at 20 RPM and 1,000 requests per day.
- Zero data retention applies to OpenCode Zen only. OpenRouter retains prompts and completions.
- The 100 trillion token capacity claim is dominated by cache reads, which cost roughly 30x less than fresh input to serve.
- A three hour eval run that burned 2.72 billion tokens would have cost between $121 and $1,977 on paid models.
- Nobody knows which lab built it. Do not architect around a guess.
- Put a hard expiry date in your config so the free window closing is an error message, not a mystery outage.
The real lesson is not about this model. It is that "free" and "unlimited" are now marketing terms with provider-specific fine print, and the only way to know what you are actually getting is to open both documentation pages and read them next to each other. That took me twenty minutes. It changed the answer completely.
Frequently Asked Questions
What is Ox Alpha and who made it?
Ox Alpha is a stealth AI model released on August 20, 2026, available on OpenRouter as stealth/ox-alpha and on OpenCode Zen as x-preview-f-free. It has a 1,048,576 token context window, a 131,072 token maximum output, and accepts text, image and video input. The lab behind it has chosen to remain anonymous during the preview, and no official reveal has happened yet.
Is Ox Alpha really free and unlimited?
It is free on both OpenRouter and OpenCode Zen for roughly one week from August 20, 2026. It is only close to unlimited on OpenCode Zen, which claims capacity for 100 trillion tokens per day. OpenRouter applies its standard free tier limits of 20 requests per minute and 1,000 requests per day for accounts with at least $10 in lifetime credits, or 50 per day below that.
Does Ox Alpha keep my prompts?
It depends on the route. OpenCode's Zen documentation states that Ox Alpha Free's provider follows a zero retention policy and does not use your data for model training. OpenRouter's model page states that prompts and completions are retained by the provider, though not used for training. If retention matters for your work, use OpenCode Zen.
Is Ox Alpha safe to use for security research?
It has notably weak refusal behavior, which is why security researchers like it for exploitation reasoning. The risk is not the model, it is the counterparty. The provider is anonymous, and on the OpenRouter route your prompts are retained. Do not paste real infrastructure details, client data or live credentials into it on either route.
What happens to Ox Alpha after the free week ends?
The most likely outcomes are a public reveal with real pricing, or the endpoint being removed entirely. The precedent is Sonoma Dusk Alpha and Sonoma Sky Alpha, which ran free on OpenRouter in September 2025 and were revealed as early builds of xAI's Grok 4 Fast. Write a fallback model into your config now so a removed endpoint fails with a clear message.
Grabbing free tokens this week? Set up your batch runs properly with the cron generator, and if you are wrangling model responses into structured data, this free formatter will save you the squinting.
Written by Adithya Guttha, Founder of Art of Code.