Union Alpha burned 2 billion tokens on day one — and nobody claims it
OpenRouter's newest stealth model has no name, no model card, and no published benchmark table — only a free week, a 262K context window, and a lot of curious developers.
A model nobody will put a name to burned roughly 2 billion tokens in its first day on OpenRouter, and more than 100 billion in total since, according to the platform's own telemetry. Union Alpha landed on the router on September 16 as a stealth listing: text and image input, a 262,144-token context window, up to 131,072 tokens of output, tool calling and JSON output, priced at zero for a limited preview. OpenRouter's model card is explicit that the model is "developed and operated by a third-party provider who has chosen to remain anonymous," and that the router is not its developer, owner, or provider. OpenCode, which wired the model into its terminal coding agent, says the serving route carries zero data retention and will not train on user prompts.
What Union Alpha does not have is the thing everyone wants to check first. There is no parameter count, no architecture disclosure, no knowledge cutoff, no weights, and no benchmark table on the listing. The phrase "frontier-level performance" appears in the description, but nothing on the card backs it. That gap is being filled by users, in public, and the results are mixed in the way anonymous model launches usually are. One tester reported Union Alpha scoring 74 on DeepSWE, matching GPT-6 Astra, with Terminal-Bench v4 near 50 percent at a cost of roughly $1.50 to $2 per task once the free window closes. Another ran a 157-case terminal suite at 136/157 with zero errors and called reasoning and tool use perfect while faulting code writing. Against that: complaints about throughput being extremely slow, a report of the model stalling in a thinking state inside OpenCode, and skeptics who note that one clean web app in a single shot says nothing about surviving a large repository for forty minutes of tool calls.
The guessing is the point of the exercise. Candidates circulating on X and Reddit range from an unreleased Claude Opus refresh to Mistral's next flagship to a low-cost sibling of OpenAI's GPT-6 — one tester says a system prompt asking it to reveal its developer returned "ChatGPT" and "OpenAI," which is evidence of exactly nothing. This is the same pattern we covered last month — Mystery model Ox Alpha tops GPT-5.6 — sleuths say it's Zhipu's GLM — and that one ended with the name coming out (Z.ai's GLM-5.3-Flash). What is different here is the volume: the previous stealth run was a curiosity, this one is processing tokens at production scale inside real coding agents.
Baseten's research arm is building an open-weight safety standard with Hugging Face and Goodfire, and it intends to run the monitoring inside its own inference platform. Base Labs says it will develop and publish methods for training open models with stronger safety properties and for monitoring them once deployed; Baseten plans to fold the output directly into its managed serving layer, Hugging Face brings distribution across the open-weight ecosystem, and Goodfire — which raised $150 million at a $1.25 billion valuation in February — brings interpretability. The trio is framing the work as a standard that is designed into training and deployment rather than bolted on later, and Baseten has issued an open call for outside contributors. The companies did not disclose how the partnership works technically, which is the part that would tell you whether this is infrastructure or positioning.
Pinterest has a new AI feature that redecorates your room, and its real purpose is to close the gap between saving a pin and buying the thing. Restyle, in beta in the US and Canada, lets a user photograph a space and then prompt edits — swap furniture, change the lighting, test styles like "bohemian" or "industrial" — or tap individual items in the image to replace or erase them, with saved Pinterest items droppable into the scene. Restyling rooms is already a crowded consumer use case for image models, so the feature is not novel; what is novel is owning it, because seeing a saved product inside your own room is the shortest path from board to checkout. Pinterest says Restyle stays an early preview before a broader rollout next month.
What to watch: whether the promised DeepSWE and Terminal-Bench numbers hold up under a harness someone publishes, and whether OpenRouter names Union Alpha's provider when the free week ends.
If a free model is good enough to run your coding agent for a week, does the developer's identity still matter to you? Tell us in the comments.
