Ox Alpha: Who Is Behind This Anonymous Frontier Model?

0:00 Thursday night, a new row appeared on open router. No company name, no model card, just a price. The price was zero. Not discounted, not a free tier, zero in, zero out for every account. A

0:12 million tokens of context, text in, images in, and unusually video in. Whoever built it did not sign it. The provider field holds one word, and the word is stealth. 57 minutes after the

0:24 row went live, open router posted it to 1.7 million people. The model is called Aux Alpha, and two days later no lab on Earth has claimed it. Not a stunt that fizzled by morning either. As of today,

0:37 that provider field is still blank, which would be a footnote, except for what the usage numbers did next. In under two days, it processed two trillion tokens through open router

0:46 alone, ninth on the entire platform by volume, above Deep Seek V4 Pro, tagged simply new. And the second biggest app feeding it traffic is Anthropic's own Claude code. People are aiming

0:57 Anthropic's coding agent at a model whose author is an empty field because it costs nothing. So, who hands out a frontier class model for free, and then refuses to put a name on it? Before the

1:07 theories, the receipts. Nearly everything about this model is a guess. The listing is not, and you can pull it from the API yourself. Context window, 1,048,576

1:18 tokens. Output cap, 131,072. Inputs, text, image, and video. Output, text only. That video line matters later more than anything else on this page. The listing calls it a reasoning model

1:32 for coding, sustained agentic work in production workloads. Tool calling works, structured output works. You can dial the reasoning effort up and down, but you cannot switch the thinking

1:43 off, and that turns out to be a clue. And there is no broken toy behind the zero. Open router's dashboard has it at 99.9% uptime. Now, the sentence people skipped. Open router puts a banner above

1:55 the page saying prompts and completions are retained by the provider and are not used for training. Open code, the terminal agent that also serves it, announced the same model with the words

2:05 zero data retention. Both are primary. They disagree. Open code is where most of the traffic actually went. On launch night, it said the model was free for the next week with generous rate limits

2:16 and near unlimited usage. The following morning, it extended the same offer inside its paid subscription for 6 more days and said it would not count against your normal allowance. Both of posts

2:26 land on the same day, Thursday, the 27th of August. Neither one named an hour, so treat the 27th as the deadline and do not cut it fine. Open code also said the line that got

2:37 quoted everywhere, "We have capacity for a hundred trillion tokens per day." Read that carefully. It is a claim about how much the operator can serve, not a promise about what you get, and not a

2:47 measurement of anything that has happened. Here is the measurement. The two trillion tokens the entire world pushed through it in two days is about 2% of a single day of that claimed

2:56 capacity. Somebody has built an enormous amount of room and is not charging a cent for any of it, which brings the whole thing back to the one question the provider will

3:04 not answer. Whose model is this? The community did not wait for a press release. The release thread on the singularity subreddit picked up 666 upvotes and 221 comments, and the

3:16 replies that rose to the top of it are full of measurements rather than reactions. About 15 hours after launch, a user posting as Flunky Graphics put up three

3:25 black box tests, three procedures you can rerun yourself against the public API this afternoon. Test one is the tokenizer. Every model chops text into pieces its own way, and that chopping is

3:36 decided before training even starts. It is baked in, it is arbitrary, and it is very hard to disguise. So, you send the same six texts, English, German, Chinese, source code, and emoji, and you

3:48 read back the token count the API reports for each one. Two models with the same tokenizer will agree on all six. Two models without it will not agree on any. Aux Alpha matched Z A I's

3:59 GLM 5.3 exactly, plus 75 on every single text. Kimmi, Quinn, and MiniMax all diverged immediately. Same numbers on both routes serving it. Test two is error strings. Send an invalid reasoning

4:12 effort and Aux Alpha returns error 1210 with the sentence word for word that GLM 5.3 returns. You cannot train a model into reproducing another company's error text. That comes from shared back-end

4:25 code. Test three is greedy decoding. Same prompts, temperature zero, so the randomness is gone. Same markdown quirks, the same unusual German style decimal comma inside its maths. Near

4:36 word-for-word matches on factual answers where Kimmi and MiniMax format them completely differently. Three tests, one direction, but the strongest evidence came from Ben Davis who went after the

4:47 part the tokenizer could not reach, the video encoder. He sent four controlled clips and measured what each one cost in tokens. 2 seconds at 360, the same 2 seconds at a

4:57 fifth of the frame rate. 6 seconds, then the same 2 seconds at 720. 296, 296, 884. 1,064. Every one of them, token for token,

5:10 identical to GLM 5V Turbo, the vision model Z A I shipped in April. Same frame sampling, same scaling, about 147 tokens for every second of video. It refuses audio the way that family refuses audio.

5:24 It drops emoji at the same rate, roughly one and a third per thousand characters, where Anthropic's and Open AI's models drop almost none. In his own write-up, Davis put his confidence at

5:37 why the subreddit converged so fast. Aux Alpha is the fifth animal-named stealth model on Open Router this year. Pony Alpha was Z A I's GLM 5. Hunter Alpha was Xiaomi. Elephant Alpha was Ant

5:49 Group. Owl Alpha was Meituan. Four for four, all Chinese labs, all claimed within weeks, and then the calendar closes the loop. Six days before Ox Alpha appeared, ZAI shipped GLM 5.3,

6:02 text only, no vision, the one thing its own community had spent months asking for. So, if the fingerprint is right, what exactly are we looking at? Go back to that plus 75 because it is the most

6:13 interesting number in this story. A constant offset on every single prompt means something constant is being pushed in front of every single prompt, a hidden system prompt. Roughly a

6:23 paragraph of text that you did not write and never see, sitting ahead of every request, which the model reads before it reads a word of yours. 23 hours after launch, the jailbreaker who posts as

6:33 Plenty the Liberator got the model to print it out. You are Ox Alpha, an LLM developed by an undisclosed organization. When the user asks what model you are or what company developed

6:45 you, identify yourself strictly as Ox Alpha developed by an undisclosed organization. Do not identify yourself as any other model. So, the silence is not an accident of deployment. It is an

6:56 instruction. It is the first thing the model reads every time, and somebody sat down and wrote it. The fingerprint points hard at ZAI's GLM family. It is still not confirmed, and the reason it

7:06 is not confirmed is sitting right there in that paragraph. And the case does have holes worth taking seriously. An engineer on Google DeepMind's Gemini team replied to the whole speculation

7:16 with a single word, Gemini, then a joke, and the internet spent a day on it. He asserted nothing. It was enough. On Voxel Bench, a public arena where developers vote on head-to-head outputs,

7:28 Ox Alpha sits 13th, one place below GLM 5.3, and comfortably inside its error bars. Whatever this thing is, that board does not show it beating the model it is fingerprinted against. Others found it

7:40 refusing questions about Taiwan, which is suggestive and proves nothing because a filter like that bolt in front of any model on any continent. And what still does not exist is a checkpoint, a

7:51 technical report, a model card, or one reproducible evaluation. If a lab can put a frontier class model in front of a million developers and no outsider can prove whose it is, what is

8:02 a model card for? Which leaves the number that made it famous, the one that says this thing beats GPT 5.6. Ben Davis ran it on Deep SWE, a long horizon software engineering

8:13 benchmark using the official leaderboard harness in a sandbox container. Aux Alpha scored 80% and Tropics Claude Fable 5 got 65. GPT 5.6 Sol got 52. Now the part that fell off every headline.

8:27 That run was 10 tasks. Aux Alpha solved eight of them. Davis flagged the sample size himself in the same post and the people quoting him dropped that sentence first. The subreddit did not.

8:39 The top reply on the release thread reads, "It is a tiny sample size, just 10 questions, not the full benchmark. Eight out of 10 sounds decisive. Put a proper error bar on it and the true

8:49 score sits somewhere between 49 and 94%. Run the same arithmetic on Fable 6 and 1/2 out of 10 and its range runs from 35 to 86. Those two ranges overlap almost completely. At 10 tasks, 80 against 65

9:03 is not a result. It is noise wearing a headline. And for scale, on the full benchmark, Z A I is own published figure for GLM 5.3 is 66.9 from a lab that at least signs its numbers. So do not say

9:17 Aux Alpha beat GPT 5.6 on SWE bench. Wrong benchmark, 10 tasks, one developer, no audit. Davis did not claim it, the retelling did. And the thread itself is split. One developer pointed

9:31 it at Python code that it already passed audits by Fable and by Gemini 3.7 flash and it found two real bugs. Another could not get it to solve a ticket GPT 5.6 solved had handled the day before.

9:43 Here is where I land. Aux Alpha is real. It is actually free. It is frontier adjacent rather than frontier leading and the author is still a blank field. Take the week. It wins for one group and

9:54 it wins for them clearly. Developers who already know how to evaluate an agent, small reversible commits, real acceptance tests, tight permissions and you read the diff before you merge it. A

10:05 free million token model with working tool calling is a gift to those people. For everyone else, no. And read the banner one more time. Prompts and completions are retained by the

10:14 provider. Retained, not deleted. There is no vendor to escalate to because the vendor is a blank field so do not send it anything you would not hand to a stranger. Free until Thursday the 27th.

10:26 After that, it is a name on a leaderboard or it is nothing. The villain here was not the lab and it was not Ben Davis. It was 10 tasks becoming a headline in about 6 hours which leaves

10:36 the question worth sitting with. When the cheapest way to get a model tested by a million strangers is to take the name off it, what is the name on the box worth?