A free model with a million token context window is the kind of thing that makes developers abandon calendar invites, lunch, and basic hydration. Ox Alpha arrived looking like someone left a frontier style coding model unattended on OpenRouter with the meter set to zero. Lovely. Also, the listing reportedly retains prompts and completions through a provider nobody has identified, which is less lovely and more like finding a raccoon in your CI pipeline wearing a badge that says trust me. ## The free lunch has a provider attached According to AI Catchup, Ox Alpha appeared on OpenRouter and OpenCode on August 20 with no named developer, a 1,048,576-token context window, text, image, and video input, and $0 pricing for input and output. Coursiv reports the same August 20, 2026 appearance and says the model is free for roughly one week, while AI Catchup adds that OpenCode described usage as near unlimited and not counting against Go plans. The Next Web supplies the eyebrow raiser: the OpenRouter listing says prompts and completions are retained by an unidentified provider. That is the whole trade in miniature: absurdly generous capability, unclear provenance, and data handling that should make anyone with customer logs sit up straighter. This does not mean Ox Alpha is bad. It means Ox Alpha is not a vending machine where free tokens fall out of the sky because the cloud loves you personally. Someone is running the model, someone is paying for inference, and someone has terms attached, even if the logo remains a tasteful cloud of fog. ## A million tokens is not a trust policy Local AI Zone reports that Ox Alpha achieved 80% Pass@1 on the DeepSWE coding benchmark, while noting that DeepSWE is a different benchmark than SWE-bench Verified. That is an important caveat because benchmark comparisons are already a haunted corn maze, and mixing suites turns the maze into performance art. The same analysis describes Ox Alpha as having a 1,048,576-token context window and multimodal support, matching the broader reports from AI Catchup and Coursiv. Coursiv says nobody has claimed the model yet and that community fingerprinting points toward a Chinese lab, with Z.ai’s GLM family as the leading theory. Treat that as a theory, not a vendor record you can staple to a procurement packet. Model provenance matters because debugging an incident later with the sentence our production data went to mystery lab, probably is how compliance teams learn to levitate. ## How to test it without feeding it the crown jewels AI Catchup says OpenRouter describes Ox Alpha as a model for coding, sustained agentic work, and production workloads, which is exactly the kind of phrase that tempts teams to route real tickets through it by Friday. Resist the raccoon. Use it for synthetic repositories, public code, throwaway agent loops, long context stress tests, and benchmark tasks where the input is not proprietary, regulated, or customer linked. The Next Web’s reported retention detail should be the gating factor for serious workloads, not the $0 price. Before routing production traffic, builders should verify provider identity, retention duration, opt out options, logging behavior, and whether prompts and completions can be used beyond serving the request. If that information is absent, the safe default is simple: no secrets, no customer data, no private source, no internal incident reports. Yes, this makes the free candy less exciting. So does reading the ingredients, and yet here we are, still alive. ## The real lesson for model routers Coursiv frames Ox Alpha as a rare chance to test a possible frontier model at zero cost while keeping sensitive data out, and that is the sane lane. The practical lesson is bigger than this one stealth release: model selection is no longer just leaderboard plus price per token. It is capability, provenance, privacy, retention, jurisdiction, fallback behavior, and whether your legal team can identify the counterparty without using a corkboard and red string. Watch what OpenRouter, OpenCode, and the eventual provider disclose next: who built it, what happens after the free window, and whether data retention terms become clearer. If you are a builder, Ox Alpha is worth testing with clean inputs because long context coding models are genuinely useful when they work. Just do not confuse free inference with free risk. The model may read a million tokens, but your procurement checklist only needs to read one word first: who. ## Sources - Free 1M-Context Coding Model on OpenRouter and OpenCode

Sources