On August 20, a new line appeared in OpenRouter's model list: stealth/ox-alpha. The vendor field reads "Stealth," and the page carries just one line of explanation: the model is developed and operated by a third party that has chosen to remain anonymous. The specs, though, are fully public — a 1,048,576-token context window, up to 131,072 tokens of output per call, text/image/video input, function calling, and JSON output support. The price column shows two zeros: $0 per million tokens for both input and output.
A price tag with two zeros
Traffic patterns are more useful here than parameters. On OpenRouter's app leaderboard, the top accounts sending work to Ox Alpha are Hermes Agent, Claude Code, and omp — all agent shells, not a single chat window among them. The page lists a P50 latency of 6.70 seconds, throughput of 20 tokens per second, three-day availability of 99.50%, and uptime of 99.99%. That speed would draw complaints in a chat product; in an overnight coding task, nobody notices.
Free looks more like trading inference cost for trajectories. A million-token context plus video input, run through repeated tool calls in an agent loop, is the most expensive request type there is. The vendor is footing that bill to collect something specific: the full record of a model calling tools fifty times in a row inside a real repository, failing, retrying, and finally getting the tests to pass green. That kind of data barely exists in public corpora and can't be manufactured by gaming benchmark leaderboards. The roughly 4.45% tool-call error rate shown on the page can only be measured accurately by feeding it real projects from real people.
A rough estimate: agent shells routinely burn hundreds of thousands of tokens on a single task. A developer who free-rides for a week saves the vendor the cost of a labeling job — one that comes with real production context and real failure cases built in.
The fingerprint points to Zhipu
Two days after launch, the community has already narrowed the field considerably. An error leaked a piece of the server-side stack trace, revealing the path com.wd.paas.api.domain.v4.chat.ChatCompletionRequest, which matches the endpoint naming in Zhipu AI's public documentation. Error code 1214 reproduces on GLM-series models hosted on OpenRouter, but not on the same-named model hosted by DeepInfra. A tokenizer probe run across 30 test groups — spanning multiple writing systems, emoji, code, and SQL — matched GLM's tokenization in all 30 cases. The person who compiled the comparison put their own confidence at 0.98, pointing to a variant of GLM-5.3. A secondary hypothesis points to Xiaomi's MiMo team. All of this remains outside inference for now; no party has come forward to claim the model.
Benchmark results are still scattered and thin. A developer named Ben Davis measured 80% on DeepSWE; in the same round, Fable scored 65% and GPT-5.6 Sol scored 52%. A single tester's single-round results prove nothing on their own, but they line up with the traffic pattern: the accounts showing up are coding agents.
Anonymous launches are becoming a release strategy
OpenRouter has seen anonymous models before. Two earlier models, code-named Hunter and Healer, followed the same path: free for a stretch, later claimed by Xiaomi's MiMo team, then given official pricing. GLM-5 and MiMo-V2-Pro arrived the same way.
The pattern is consistent enough to read as a deliberate release strategy — collect data and build a reputation during the anonymous window, while dodging the pile-on comparisons that hit every model on its launch day; once leaderboard rank and developer goodwill are secured, drop the mask and switch to paid pricing.
For Chinese vendors, there's an extra payoff. When a model launches carrying the label of a specific Chinese lab, overseas developers' first reaction often lands on where it came from rather than what it can do. Unbranded, the judgment has nowhere to land but the code itself. By the time the identity comes out, the favorable reviews are already sitting in forum threads and tweets.
The cost of "free" is spelled out on the page. OpenRouter's data policy states that prompts and outputs are retained by the provider, with only a promise not to use them for training; OpenCode's direct route, meanwhile, advertises zero data retention. The two statements don't agree with each other across the same request chain. Running an open-source project through it carries little risk — but before dropping an entire private company repository into that 1-million-token window, that fine print is worth reading first.
OpenRouter describes the preview as a time-limited experiment without giving an end date. When the mask comes off, and what price follows, nobody is saying yet.
Sources: OpenRouter's model page and provider data-policy notes, CocoLoop, developer-community fingerprinting records and public benchmark posts; context window, max output, latency, and throughput figures follow OpenRouter's model page, and the DeepSWE score is from a single developer's single-round test.