Yesterday was Nvidia AVO on the public ARC set and a six billion Poolside factory license. This morning the unsigned model is the story.

Someone parked a frontier checkpoint on OpenRouter and declined to sign it.

Lead: Ox Alpha is a preview, not a leaderboard

OpenRouter listed stealth/ox-alpha on August 20. The card is unusually specific for a ghost: 1,048,576 token context, 131,072 max output, text plus image plus video in, text out, tools and JSON out, zero posted price. OpenRouter says it only routes. The developer is “a third-party provider who has chosen to remain anonymous during this preview.” Prompts and completions are retained by that provider and “are not used for training.”

Business Insider is the clean mainstream write-up. Stripe CEO Patrick Collison tried it and called it “very impressive.” OpenCode said the free week runs through about August 27, with a provider capacity claim of 100 trillion tokens per day. Treat that number as marketing until a lab owns it.

The viral coding score is the part you should not paste into Slack without the footnote. A 10-task community DeepSWE sample circulated at 80 percent against roughly 65 for Claude Fable and 52 for GPT-5.6 Sol. Coursiv and later write-ups are explicit: official DeepSWE is 113 long-horizon tasks. A later full-set community run landed closer to 63 percent. That is still a frontier-band coding model. It is not an audited beat.

OpenRouter’s own app graph is the more honest signal. After three days the top senders are Hermes Agent at 1.59 trillion tokens, Claude Code at 800 billion, and DeepSeek Harness at 674 billion. People are not theorizing. They are dumping repos into a free 1M window.

Bold insight: the product is not the checkpoint. The product is a week of unpaid eval plus a named custody problem. You cannot name the company holding the prompts, and the card still says they keep them.

If you are actually going to throw a private tree at it, do it from a scratch box, not the laptop with the production keys. A cheap used ThinkPad and an external SSD are the boring way to keep the experiment off the machine that holds the real secrets.

Secondary: the stack trace has a package path

TechTimes, this morning, is the day-after forensic note. Community probes did the usual tokenizer matching, then got less cute.

This is the fifth stealth drop in about six months. The last four — Z.ai’s own GLM-5, Xiaomi’s MiMo-V2-Pro, Ant’s Ling-2.6-flash, Meituan’s LongCat-2.0 — were claimed after the free week. The playbook is now a launch ritual: hide the logo, buy honest eval, convert a slice of the traffic when the name drops.

Nobody from Zhipu or OpenRouter has confirmed it. Andrew Curran noted Saturday that the GLM theory was hotter Friday night and fuzzier by morning, and a competing tokenizer read even floated Microsoft MAI. Unsigned is the correct headline.

The custody line is the one that should change what you paste. The model page says the provider retains prompts and does not train on them. Platform stealth terms and the per-model notice do not read the same way in every write-up. Until a named lab publishes a system card, treat Ox Alpha like a public eval cluster. Throw synthetic bugs at it. Do not throw customer code.

If you want the same long-context habit without an anonymous keeper, a workstation GPU plus an NVMe enclosure is still how you learn what a million-token window actually costs when the invoice has a logo.

Tertiary: Sol got cheaper. The harness market got louder.

Same 48-hour window, different layer. OpenAI cut GPT-5.6 Sol from 5 / 30 to 4 / 20 per million input / output tokens for three months, through at least November 21. Cached input follows to 0.40. Pro / Plus / Business seat prices did not move. The API pricing table matches. Output is the line that matters on agent runs. A third off that number is a fleet discount, not a hobby coupon.

That cut only makes sense next to the free ghost. Sol is still the named workhorse. Ox Alpha is the unnamed stress test. If your weekend loop can tolerate an anonymous keeper, the price is zero. If it cannot, Sol just got less embarrassing against Grok 4.6 and the cheap Chinese APIs.

TrueFoundry open-sourced TrueForge under MIT. VentureBeat has the vendor numbers. Same 14 Enterprise-Bench tasks: GLM-5.2 on TrueForge at 2.90 versus Opus 4.8 on Claude Managed Agents at 11.80 (about 75 percent cheaper); same Opus backbone still about 30 percent cheaper (8.50 vs 11.80). Those are the company’s benches. The architectural claim is the useful one: sandbox is a tool, not a permanent jail, and context gets compacted instead of re-sent.

Yesterday Nvidia showed a closed harness turning a 30-percent model eval into a 100 on a public game set. Today an open harness is selling cost-per-completed-task. Same week, same layer of the stack.

If you are going to run either loop overnight, a USB-C dock and a second monitor still beat a brown-out mid-eval.

Takeaway

1. Today: try Ox Alpha on throwaway tasks before August 27 if you want the free window. Read the OpenRouter card as a custody notice, not a system card.
2. Do not treat the 80 percent DeepSWE clip as a leaderboard. Ten tasks is a vibe. 63 percent on a fuller run is still interesting. Neither is official.
3. Pricing: Sol at 4 / 20 through November 21 is the named alternative when the ghost week ends.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.