What this actually is
On August 20, 2026 a model called ox-alpha appeared on OpenRouter with no announcement, no company attached to it, and a price of zero. Its listing says the provider is a third party who has chosen to remain anonymous for the length of the preview. There is no lab name, no launch post, and no pricing page.
What it does have is a specification that reads like a frontier release: a 1,048,576 token context window, up to 131,072 tokens of output in a single response, text, image and video accepted as input, and tool calling and structured outputs supported. It is aimed squarely at coding and long agent runs.
The free preview ends on August 27, 2026. After that it either gets a price or it disappears, and nobody outside the lab knows which.
The honest version of the benchmark
You have probably seen a number. Here is where it comes from, because the gap between the number and its provenance is the whole story.
- The figure in circulation is DeepSWE Pass@1: ox-alpha 80 percent, Claude Fable 5 at 65, GPT-5.6 at 52. That is a real result and the ordering has held up across several independent write-ups.
- It came from a ten task run by one developer, not from an audited leaderboard. Ten tasks is a small sample. A single task swinging either way moves the headline by ten points.
- OpenRouter itself publishes no scores for this model at all. No intelligence index, no coding leaderboard number, no agentic score. Every figure you have read is community reported.
- Other numbers are circulating too, including an 88 percent SWE-bench claim, and they do not reconcile with each other. Treat any single figure as one person's run until somebody audits it.
So: promising, and worth an evening. Not something to move a production stack onto this week.
The same caution applies to the capacity claim. The operator states 100 trillion tokens per day of serving capacity and near unlimited rate limits. That is the operator's own claim, not a measurement, and 100 trillion a day is in the range only the largest providers claim across every model and customer combined.
Free, for your email
Get the full setup steps + the copy-paste setup prompt
The rest of this guide (5 more sections) opens right here. The setup prompt, one paste into Claude that sets it all up, comes by email only. Use your real address.
Unlocks: The setup, in two lines · One warning worth reading before you commit an afternoon · Where else it runs · What to actually build with a million token window · The short version
50,000+ follow where these guides come from. No spam, reply “unsubscribe” any time.