Ox Alpha: The Mystery Model That Took Over OpenRouter

TL;DR: Ox Alpha is a free, anonymous reasoning model on OpenRouter with a 1,048,576-token context window, 131,072-token maximum output, text/image/video input, and tool calling.^[1][2] OpenRouter's daily activity feed shows usage rising from 27,032,893 requests and 1.9909 trillion prompt-plus-completion tokens on August 21 to 70,334,611 requests and 5.9296 trillion tokens on August 24.^[3] It ranked first among 558 models in OpenRouter's rolling-week popularity sort in our August 25 snapshot, but neither traffic nor community fingerprints reveal who built it.^[4][5]

Here is the simple version. Imagine a car company parks an extremely capable test car outside a racetrack. Anyone can drive it for free. The badges are covered, the engine bay is locked, and the company refuses to say who made it. You can measure how many laps people drive and how the car behaves. You cannot prove who built it from the exhaust note.

That is Ox Alpha, officially spelled with the letter O, although people often write “0xAlpha” online. OpenRouter added the endpoint stealth/ox-alpha on August 20, 2026. It says an unnamed third party develops and operates the model, while OpenRouter only routes requests.^[1][2]

What does it actually do? It is built for coding and long agent runs. It can accept text, images, and video, produce text, call software tools, return JSON, and keep a huge amount of working material in one request. The maker is the mystery. The product and the demand are visible.

The Demand: 70.3 Million Requests In One Day

The cleanest measure of Ox Alpha's rise is OpenRouter's own daily model-activity feed. The endpoint is official but undocumented, so its figures should be treated as a time-stamped platform snapshot rather than audited financial reporting.^[3]

On August 21, its first complete day, Ox Alpha handled 27,032,893 requests, 1,966,507,869,619 prompt tokens, and 24,405,377,089 completion tokens. By August 24, it handled 70,334,611 requests, 5,835,184,092,873 prompt tokens, and 94,402,126,602 completion tokens.^[3]

That is a 160.2% increase in daily requests and a 197.8% increase in combined prompt-plus-completion tokens in three days. Across August 20 through 24, the feed records 216,084,951 requests and 17.4977 trillion combined prompt-plus-completion tokens. August 20 was only a partial day because OpenRouter's catalog record was created at 20:04:55 UTC.^[2][3]

The popularity ranking reinforces the daily series. OpenRouter documents that most-popular and top-weekly sort models by tokens processed during the last week.^[4] At 03:15 UTC on August 25, its models API placed Ox Alpha first among 558 returned models.^[5]

Let's be clear: this is not a quality leaderboard. A free model with a million-token window has an enormous mechanical advantage in a ranking based on processed tokens. Long coding-agent sessions also reread context, ingest tool output, retry failed actions, and cache large prefixes. Usage proves attention and workflow placement. It does not prove that Ox Alpha is the best model.

The Product: A Million-Token Workspace For Agents

Ox Alpha's appeal is easy to understand. OpenRouter lists 1,048,576 tokens of context and 131,072 tokens of maximum output. Inputs can include text, images, and video; output is text. The endpoint accepts tools, tool_choice, reasoning controls, and response_format, although JSON output does not enforce a supplied schema.^[1][2]

OpenRouter's launch post on X called it a model for coding, sustained agents, and production, with one-million-token context and multimodal input.^[13] That is the platform's positioning, not independent proof that Ox Alpha belongs in a frontier tier.

In ELI5 terms, the context window is the size of the model's desk. A normal desk may hold the file you are editing and a few notes. A million-token desk can hold a large repository, logs, issue history, design documents, screenshots, and a long trail of tool results at once. That does not make the worker wise. It gives the worker more material to look at before deciding what to do.

OpenRouter's page displayed 3.88 seconds P50 latency, 25 tokens per second P50 throughput, 100.00% three-day uptime, and 99.35% availability in our snapshot.^[1] Those are routed deployment signals, not competitor benchmarks. Hardware, precision, prompt and output length, decoding settings, speculative acceptance, batch size, concurrency, time to first token, tail latency, and the agent harness are not normalized against another model.

The real story isn't that Ox Alpha offers a specification no one has seen before. It is that the model combines the right ingredients for expensive agent loops, then prices every token at zero while developers decide what to integrate.

The Identity: Fingerprints Narrow The Family, Not The Owner

No named lab had claimed Ox Alpha when this article was published. OpenRouter explicitly says it is not the developer, owner, or provider. That is the highest-confidence identity fact available.^[1]

The leading community theory points toward GLM-family lineage and a Z.ai-compatible serving stack. The open-source modelprint project reports that Ox Alpha matched z-ai/glm-5.3 on 6 of 9 infrastructure probes and all 4 normalized tokenizer probes. The next GLM candidate matched 5 of 9; non-GLM candidates matched 2 of 9 or fewer in that published run.^[9]

That is real evidence. It is not ownership proof. modelprint itself explains that a router can expose router-generated behavior, two models can share infrastructure, and a familiar tokenizer can identify lineage without identifying an exact checkpoint. Behavioral questions, political answers, screenshots, and model self-identification are weaker still because system prompts can shape them.

Z.ai's published GLM-5.3 interface has notable similarities: a one-million-token context, 128K output, mandatory reasoning, low/high/max effort settings, temperature 1, and top-p 0.95.^[10] It also has a notable conflict. Z.ai documents public GLM-5.3 as text-only, while Ox Alpha advertises image and video input. The defensible conclusion is “consistent with GLM-family lineage and Z.ai-compatible serving,” not “made by Z.ai.”

The Benchmarks: One Full Run Beats A Viral Ten-Task Slice

Ox Alpha has no provider-published model card, benchmark paper, or audited public leaderboard. That makes the early social-media numbers especially easy to misuse.

One viral DeepSWE post reported 8 of 10 tasks and explicitly warned that the tiny subset could have substantial variance.^[14] Ten tasks are less than 9% of the official 113-task benchmark, and a handful of successes or failures can swing the percentage dramatically.^[11] The same tester later corrected the public impression with a roughly 63% result from a larger community run and said the first subset's 80% did not hold.^[15] Neither post should be used to claim Ox Alpha beat GPT, Claude, Grok, GLM, or any other system whose result used a different task count, retry budget, harness, or reasoning configuration.

A more useful self-published run evaluated all 113 DeepSWE tasks using pier 0.3.1, mini-swe-agent, a Docker backend, concurrency 4, and up to two retries. It reports 66 resolved tasks, or 58.4%, with a Wilson 95% confidence interval of 49.2% to 67.1% over 20 hours and 39 minutes.^[12]

That result is still not an independent audited leaderboard. The repository reports five timeouts and tool-call formatting failures, and one trajectory per task does not measure run-to-run variance. Its value is methodological. It tells readers exactly what ran, over the full set, with a named harness and visible artifacts.

The Fine Print: Free Is A Data-Governance Decision

OpenRouter's Ox Alpha page says the anonymous provider retains prompts and completions but does not use them for training.^[1] Retention by an unnamed operator is already enough to keep customer repositories, credentials, personal data, regulated material, and unreleased strategy off the route.

The broader OpenRouter Stealth Program EULA creates a second problem. Updated July 6, it says stealth models exist for collecting user content for training and improvement, grants OpenRouter a perpetual license to use and store that content, and permits sublicensing it to the applicable stealth provider for training, evaluation, and improvement.^[6] We found no Ox Alpha-specific supplemental terms resolving that conflict.^[7]

The model page and the program contract therefore point in different directions on training use. The responsible operating assumption is the stricter one until OpenRouter publishes a model-specific legal clarification.

OpenCode offers Ox Alpha through a separate Zen route and describes that route as zero retention with no training use.^[8] Those assurances do not automatically transfer to OpenRouter. Same apparent model, different intermediary, potentially different serving configuration and data terms.

The Strategy: Free Inference Bought A Distribution Position

The uncomfortable truth is that mystery is excellent marketing when the product is free and technically interesting. Developers run the tests, build the integrations, publish the comparisons, and promote the identity puzzle. The provider receives a real-world load test and a market map before spending its own brand equity.

OpenRouter's activity feed adds a sharper commercial clue. On August 24 it records 43,270,460 tool calls, equal to 61.5 tool calls for every 100 requests if read as separate counters from the same daily row.^[3] That does not prove 61.5% of requests called a tool because requests can contain multiple calls. It does show that agents, rather than passive chat alone, were central to the workload.

Here's the genius: a zero-dollar route removes price comparison before model comparison begins. A million-token context encourages enormous working sets. An OpenAI-compatible API lowers integration friction. The anonymous provider buys a place inside agent harnesses, then learns which workflows, prompts, context shapes, and failure modes matter at scale.

This strategy has a cost for the provider. Serving trillions of tokens is not free, even when the price displayed to users is zero. It also creates no durable promise for buyers. OpenRouter's Stealth EULA says models can disappear with or without notice.^[6] A team that replaces its production model because the preview costs nothing is not saving money. It is borrowing dependency risk.

While competitors launch named products with benchmark tables, the Ox Alpha provider was buying workflow placement through subsidized inference. The traffic curve shows that the tactic worked. The missing identity and contradictory data terms show why adoption is not the same as procurement.

The Verdict: The Usage Reveal Matters More Than The Name Reveal

Ox Alpha may eventually be claimed by Z.ai, Xiaomi, Microsoft, another lab, or a company no one has guessed. Today, none of those names is confirmed. The evidence can narrow infrastructure lineage. It cannot sign a vendor contract.

The real story isn't who wins the guessing game. It is that an anonymous provider moved from zero public identity to OpenRouter's number-one rolling-week model, with daily demand reaching 70.3 million requests and 5.93 trillion prompt-plus-completion tokens, by combining free inference, a million-token context, tool support, and distribution through a standard API.^[3][5]

Ox Alpha proved that developers will adopt capability before provenance. The next phase will test whether enterprises are willing to do the same. They should not be.

Last updated: August 25, 2026