Union Alpha Unmasked: OpenRouter's Free Stealth Model Is Pareto by Unbiased

For a few days, one entry in OpenRouter's catalog carried no name at all. Union Alpha showed up as a stealth model: free while testing lasted, able to read images and call tools, with a 262,000-token context window. Where the maker's name usually sits, the page said exactly one word: stealth. That mystery has now been resolved on the record, because OpenRouter's own model page lists Union Alpha as a stealth model that was revealed to be Pareto by Unbiased, with the live entry at openrouter.ai/unbiased/pareto. The episode deserves more than a footnote for anyone running agents: it shows how models get tested, teased and priced in public without a launch event, and it is a small lesson in cost, risk and attribution all at once. Everything below reflects the sources as of September 26, 2026.

OpenRouter works like a shop for language models: you walk in, pick a model, and plug it into whatever tool you use, without a separate account for every provider. Hundreds of models, one login, one billing path. Next to the official listings, the catalog keeps a dedicated namespace for models under assumed names, in this case stealth/union-alpha. That is where an unknown publisher parked the model, priced at zero during the test phase, with a single hint about identity: the word stealth instead of a company name. For users such a listing is tempting and a blind flight at once, because capabilities can be tried while origin, data policy and post-trial costs stay open. That exact combination pulled the community onto the case within hours.

What the model could do before it had a name reads plainly from the description and the first tests: Union Alpha accepts text and images, can use tools, and is therefore built for jobs that go beyond plain chat, such as research, coding tasks or multi-step agent runs. The context window sits at 262,000 tokens, solid by current standards, though the short video by Julian Goldie notes it is not quite at frontier level. OpenRouter now describes the same model under the Pareto name as multimodal, built for research, coding and agentic workflows, with frontier-level performance across a broad range of general-purpose tasks. That phrasing is marketing until an independent measurement backs it up, but it marks the audience precisely: people who operate agents and need a model that carries work steps, not just conversation.

The sequence after the model surfaced was remarkably fast and fairly well documented, at least in the retelling of the short video: about ten minutes after the appearance, the coding tool OpenCode told its users the model would be free for a week. Twenty-seven minutes in, an OpenRouter founder posted a benchmark chart in which Union Alpha appeared as a gold star near the top of a coding benchmark, with no number printed next to it. A separate post by Alex Atallah, co-founder of OpenRouter, shows a DeepSWE result for the model and makes clear the stealth entry was actively promoted from the platform side. Then the floodgates opened: people piled in, started digging, and compared spend, latency and behavior patterns. Unmasking a model before its maker does buys attention that normally only an official launch commands.

The core of the story is one number: roughly 73 percent on a software-engineering benchmark, close to the roughly 74 percent that GPT- and Opus-class models reach there, at a cost of about $0.65 per task. These figures come from a compiled MindStudio analysis that explicitly warns there is no independent confirmation and that throttling limited testing during the stealth phase. Numbers from this phase deserve caution anyway, because throttling keeps the sample small and the task mix is unknown. The sober reading stands: a chart without a printed number is marketing until someone reproduces the task. Still, the per-task cost of about $0.65 fits the story of a model that wants to deliver frontier-adjacent work at a fraction of usual prices. Whether the number holds gets decided once the model is stably reachable outside the test window and third parties can rerun the same tasks.

Among the interpretations that circulated in the first hours, one is especially interesting because it brushes against the business model of routing platforms: a Reddit thread describes Union Alpha as a blended model that queries several language models in parallel per request and synthesizes one answer from them. That is unconfirmed, and the vendor page says nothing about it. If true, it would carry consequences: an ensemble of several models makes low per-task cost hard to explain, because someone must pay for the parallel computation. It would, however, explain why the benchmark curve looks so smooth, since blending irons out the typical outliers of single models. While the vendor stays silent, the community fills the gap, and the most plausible reading wins reach regardless of its evidence. For now this remains a community hypothesis, not a fact.

The resolution arrived in an unusually quiet way: no press conference from the maker, just the model page itself. OpenRouter now lists Union Alpha with the note that the stealth model was revealed to be Pareto by Unbiased, and links the live entry openrouter.ai/unbiased/pareto. So the attribution question has an answer: a vendor calling itself Unbiased. What Unbiased otherwise does, how large the team is, and under what conditions Pareto will be offered after the test window are not officially documented as of retrieval, and the price is missing too. That is the real trick of the stealth pattern: attribution arrives once the test phase has delivered the data, not before. For model users the takeaway is simple: a free window is the best moment to stress-test a model under load, but contracts and evaluations should wait until a name and a price sheet exist.

Three points stick when you read the case through the eyes of people who run their own agents. First: stealth testing has become a normal distribution channel, not a curiosity. Labs place models under codenames in routing catalogs to learn behavior, latency and cost under real load before an official launch sets expectations. Second: benchmark charts without printed numbers are invitations, not proof. The roughly 73 percent remain a claim until independent reruns drive the same tasks. Third: free test windows are the cheapest way to evaluate an unknown candidate inside your own agent fleet, but risk checks must not pause just because the model costs nothing. Running critical tool calls against a model whose identity can flip mid-experiment is exactly what happened here, literally. Practically: log tool calls, keep approval thresholds in place, and treat every free phase as a test, not a migration.

The case fits a pattern of this year: model reveals have moved from press conferences to routing platforms, and the community completes attribution in hours rather than weeks. For agent practice the shift matters more than the anecdote: the most interesting models of the quarter surface under codenames first, and testing them early in your own stack takes tooling and processes that survive identity changes. That is why this case sits on the Hunter radar rail: agents operate through routers like OpenRouter, and every new entry there changes the math, from per-task price to risk checking.