Ox Alpha is a reasoning model that appeared on OpenRouter on August 20 with input and output free of charge, a context window of 1,048,576 tokens, and a maximum single output of 131,072 tokensCITE:E1. OpenRouter's official X account introduced Ox Alpha the same day as a "frontier model" built for efficient coding, sustained agentic work, and real-world production use, describing a 1M-token context window with text, image, and video inputCITE:E15. Ox Alpha's own listing claims it can process up to 100 trillion tokens per day, offered free during the trialCITE:E13. Separately, its context is described as multimodal — spanning text, audio, and video — at more than 1 million tokens, paired with the 131,072-token output cap and a one-week free window at up to 100 trillion tokens a dayCITE:E14. OpenCode's documentation lists the same model as "Ox Alpha Free," noting the free access is time-limitedCITE:E3.
Ox Alpha reached No. 4 on OpenRouter's weekly usage chart by August 22, processing 6.54 trillion tokens that week with its trend column marked "new"CITE:E4.
| Rank | Model | Weekly tokens |
|---|---|---|
| 1 | DeepSeek V4 Flash 0731 | 11.6 trillion |
| 2 | MiMo-V2.5 (Xiaomi) | 9.14 trillion |
| 3 | Hy3 (Tencent) | 8.56 trillion |
| 4 | Ox Alpha | 6.54 trillion |
The same weekly chart placed DeepSeek V4 Flash 0731 first at 11.6 trillion tokens, Xiaomi's MiMo-V2.5 second at 9.14 trillion tokens, and Tencent's Hy3 third at 8.56 trillion tokensCITE:E5. TechCrunch reported that Stripe CEO Patrick Collison, whose company had just announced its acquisition of OpenRouter, said trying Ox Alpha was "very impressive"CITE:E6. Collison posted the same assessment directly on X, writing "this is very impressive"CITE:E16.
OpenRouter defines Ox Alpha on its official model page as a stealth model "developed and operated by a third party that has chosen to remain anonymous during the preview period," adding that OpenRouter only routes requests to it and is not its developer, owner, or providerCITE:E2. OpenCode's documentation likewise lists it only as "Ox Alpha Free" with a time-limited free tag, without naming a developerCITE:E3. As of August 24, no party has stepped forward to confirm Ox Alpha's provider identityCITE:E12.
OpenRouter has previously revealed a stealth model's identity itself: its official X account announced in February that another anonymous preview model, Pony Alpha, was in fact Zhipu's GLM-5, a 744-billion-parameter base model, disclosed once its anonymity period endedCITE:E9.
OpenRouter and OpenCode publish conflicting statements about how Ox Alpha handles user data: OpenRouter's Ox Alpha page states that prompts and completions are "retained by the provider but not used for training," governed by the Stealth Model Terms, while OpenCode's documentation says the same model's provider follows a "zero-retention policy" and will not use user data for trainingCITE:E10. OpenCode's documentation also lists several exceptions to its general policy: for Big Pickle, MiMo-V2.5 Free, and Hy3 Free, data collected during the free period "may be used to improve the model"; requests routed to OpenAI and Anthropic APIs are retained for 30 days under each company's own data policy; and Muse Spark 1.2 Contributor Free offers steeply discounted token pricing in exchange for rights to use prompts and completions to train future Meta modelsCITE:E11.
The modelprint project, run by developer unclecode, ran nine probe tests plus tokenizer comparisons against candidate models and found partial matches rather than a confirmed identityCITE:E7.
| Reference model | Probe match (9 tests) | Tokenizer match (4 tests) |
|---|---|---|
| GLM-5.3 (Zhipu) | 6/9 | 4/4 |
| GLM-4.7-flash | 5/9 | 4/4 |
| GPT-5.6-luna (OpenAI) | 2/9 | 2/4 |
Ox Alpha matched GLM-5.3 on 6 of 9 probes and all 4 of 4 tokenizer tests, matched GLM-4.7-flash on 5 of 9 probes with the same 4-of-4 tokenizer result, and matched OpenAI's GPT-5.6-luna on only 2 of 9 probes and 2 of 4 tokenizer testsCITE:E7. unclecode cautioned on the same page that "a matching fingerprint only proves shared infrastructure, not shared identity — the same lab can run two different models on the same service stack"CITE:E8.
Developer speculation, based on the model's name, capabilities, benchmark scores, and compute resources, has named several candidates: an unpublished model from Zhipu's GLM series, a variant of DeepSeek-V4-Flash, an unreleased Grok model from Elon Musk's SpaceXAI, or an unreleased Microsoft MAI model referred to as MAI-2CITE:E17. A separate rumor holds that Ox Alpha uses cl100k_base, a byte-pair-encoding tokenizer developed by OpenAI, and argues that because Chinese AI labs' models do not use OpenAI-built tokenizers, this raises the likelihood of a Microsoft originCITE:E18. Robert Lukoszko posted a tokenization-based analysis on X stating that the cl100k_base tokenizer rules out OpenAI, Google, Anthropic, xAI, and every Chinese frontier lab, and matches only Microsoft's Phi/MAI lineage, with IBM Granite as an outside possibilityCITE:E19.
What this means: Ox Alpha's usage numbers and the tokenizer-based elimination point in different directions from unclecode's fingerprint tests — one line of evidence leans toward Zhipu's GLM family, another toward Microsoft's Phi/MAI lineage — and OpenRouter's own February disclosure of Pony Alpha as GLM-5 shows the company is willing to name a stealth model once its preview period endsCITE:E9. Until that happens for Ox Alpha, the contradictory retention policies published by OpenRouter and OpenCodeCITE:E10 mean the terms governing the 6.54-trillion-token-a-week free trialCITE:E4 remain unsettled even as usage keeps climbing.
Ox Alpha is a reasoning model that appeared on OpenRouter on August 20 with input and output free of charge, a context window of 1,048,576 tokens, and a maximum…
Ox Alpha reached No. 4 on OpenRouter's weekly usage chart by August 22, processing 6.54 trillion tokens that week with its trend column marked "new"CITE:E4.
OpenRouter defines Ox Alpha on its official model page as a stealth model "developed and operated by a third party that has chosen to remain anonymous during th…
OpenRouter has previously revealed a stealth model's identity itself: its official X account announced in February that another anonymous preview model, Pony Al…
The interesting signal here isn't the 6.54-trillion-token weekly usage or the 100-trillion-token-a-day claim by themselves — it's that two independent detection methods, unclecode's fingerprint probes and the tokenizer-based elimination cited by Robert Lukoszko, land on different candidates from the same data point (cl100k_base and probe overlap don't agree on GLM versus Microsoft's Phi/MAI lineage). That divergence is the real story: stealth-model testing tools are still immature enough that shared infrastructure, shared tokenizer, and shared benchmark behavior each tell a different partial story. The concrete thing to watch is whether OpenRouter repeats its February move — when it named Pony Alpha as Zhipu's GLM-5 once the preview period ended — for Ox Alpha, and whether that eventual disclosure lines up with the GLM-leaning fingerprint results or the Microsoft-leaning tokenizer analysis.
Global semiconductor sales hit $403.3B in Q2 2026, up 35.1% from Q1, as WSTS lifts its 2026 forecast to $1,655B on surging memory and HBM demand.
Taiwan's banking sector posted unrealized stock market gains of NT$894.9 billion at the end of June 2026, up 124% year-on-year to a historical high, while adding a net NT$489.6 billion in bonds during the first half. In June alone, banks cut Taiwan stock holdings by NT$6.4 billion but turned net bond buyers with NT$48.6 billion in purchases. Taiwan's banking, insurance, and securities sectors together held NT$5.8838 trillion in stocks and NT$32.1531 trillion in bonds.
The US Treasury doubled its 10-to-30-year bond buyback size from a maximum of $2 billion to at least $4 billion per operation, running September 9 to November 4, after the 30-year yield hit a 19-year high of 5.34% on August 18. The yield fell 9 basis points to 5.19% after the announcement, but analysts note the buyback is small next to a $32.2 trillion market.