Ox Alpha surfaced this week on OpenRouter and OpenCode as a free model, giving developers temporary access to a system with a 1,048,576-token context window, multimodal input and zero pricing for both prompts and completions. On OpenRouter, it was listed under the model ID stealth/ox-alpha, a sign that the model is being offered without a publicly confirmed builder.
That is why Ox Alpha is drawing attention now: developers are not just looking at another coding model, but at one that can take text, images and video while handling extremely long codebases in a single pass. OpenCode said the model would stay free for the next week, and it also described zero data retention and generous rate limits with capacity for 100 trillion tokens per day, a package that makes the model easy to try at scale while the window is open.
ModelsAtlas dated the release to August 20, 2026, which fits the way the model began circulating as a short-term preview rather than a fully explained public launch. The numbers matter here. A 1M-token window means a model can keep far more project context in view than standard systems, and OpenRouter also showed a 131,072-token maximum completion, leaving room for long outputs as well as long inputs.
But the presentation is ahead of the proof. Ox Alpha has been described as having strong developer-reported coding benchmark results, yet OpenRouter’s entry did not show an independent benchmark block, and DeepSWE did not list it on the public BenchSift leaderboard as of August 21. The top public rows there were still held by named models such as Claude Opus 5 and GPT 5.6 SOL, with the best pass@1 results in the low 70s, which leaves Ox Alpha in a strange place: visible enough to use, but not yet anchored by public scoring.
That gap may shape how the model is judged over the next week. If Ox Alpha keeps drawing developers because it is free and unusually roomy, the market will have time to test it on real repository work instead of relying on benchmark talk. If the benchmark picture stays incomplete, then the model’s appeal will rest less on proof than on access, and the unresolved question will be simple: who built it, and what exactly is behind the anonymous release?
