A Mysterious AI Model Just Appeared. Nobody Knows Who Built It.
A model showed up on OpenRouter with a million-token context window, video input, tool use, and a preview price of zero. It has no developer attached to it. No lab has claimed it. It's called Ox Alpha, and by the time Oscar and I sat down to record Episode 20, people were already routing real work through it.
That's the thing I couldn't let go of. Not "is it good" — the early coding results are genuinely strong. The question I kept circling is simpler: you're sending your code to an endpoint, and nobody will tell you whose endpoint it is.
Signal or Noise
Oscar and I run every story through the same filter: is this signal you should act on, or noise dressed up as news? Here's how I read them this week.
Ox Alpha, the model with no author
Free during preview. Text, images, and video in. Tool use supported. Up to 131,072 tokens back. A million-token context window. On paper that's a frontier-class spec sheet with no invoice attached.
Two things keep me from getting excited. The first is evidentiary: the coding results people are posting look good, but the sample is small and the comparisons are uneven — different harnesses, different prompts, different tasks, then a leaderboard screenshot. That's not a benchmark, that's a vibe.
The second is the one I'd actually put in a policy doc. It retains your prompts. A free preview from an unnamed operator that keeps what you send it is not free — you're paying in data, and you don't know who's collecting. If you wouldn't paste it into a stranger's web form, don't route it through Ox Alpha because the token price is zero. Signal, with the caveat doing most of the work.
NVIDIA AVO clears all 183 ARC-AGI-3 levels
NVIDIA's AVO completed every one of the 183 public ARC-AGI-3 levels, running Claude Opus 5 underneath. The number that matters to me isn't 183, it's which part did the work. The model was an off-the-shelf frontier model. The result came from the harness around it — the scaffolding that decomposes, retries, and keeps state across a long horizon.
I've been saying some version of this to clients for a year: the differentiator is drifting from the model to the system you wrap around it. This is a clean data point for that. Signal.
OpenAI pauses frontier RL training over Astra
OpenAI paused frontier reinforcement-learning training for two weeks after preliminary cyber results from Astra. Whatever you think of the company, a lab voluntarily stopping a training run because early numbers spooked them is not the industry default, and it's worth noticing when it happens.
I'd hold the applause until we see what "paused" meant in practice and what changed after. But the shape of it — measure, get an uncomfortable result, stop — is the shape you want. Signal.
The Reconstruction benchmark: 3 to 15 percent
This is the story I'd hand to anyone who thinks the research-automation problem is nearly solved. Reconstruction is a blind benchmark: can a model recover the original research idea? Seven frontier models scored roughly 3 to 15 percent.
That's not a rounding error away from human researchers. That's a different activity. These systems are extraordinary at working a problem you hand them and still poor at finding the problem worth handing over. If your AI roadmap has "generates novel research directions" on it, this is the number to put next to that line. Signal, and the most useful one this week.
DeepSeek Harness can now mount Codex and Claude Code
The rc.8 release lets DeepSeek Harness install Codex and Claude Code as subagent components. Coding agents are becoming parts you compose rather than products you pick. That's good for anyone building — and it quietly makes the supply-chain question harder, because "which agent touched this repo" now has more than one answer.
Ship It or Skip It
This is my favorite segment, because it's where Oscar and I stop reporting and start arguing about what we'd actually build.
- CI for agent systems. If the AVO result is right that the harness carries the outcome, then the harness is the thing under test — and almost nobody tests it. You'd want the agent scaffolding run against fixed tasks on every change, with regressions failing the build like any other. This one follows directly from the week's news.
- An R&D idea tournament. A bracket that runs candidate research directions against each other and promotes the survivors. Interesting, and the Reconstruction numbers are exactly the reason it's hard: the models are weakest at the generative end of research, which is the end a tournament needs to be strong.
Where I landed
Three things I took away after we stopped recording:
- An anonymous free model is a data-handling decision before it's a technical one. Read the retention terms, then decide.
- When a frontier model clears a benchmark inside somebody's harness, credit the harness. That's where the engineering lives.
- Three to fifteen percent on recovering a research idea is the most honest number in this episode.
The through-line, if there is one: the interesting work this week wasn't a new model. It was the systems built around existing models — and the one benchmark that showed us where those systems still can't go.
Watch or listen to the full conversation below, and tell me whether you'd route production traffic through a model nobody has claimed.
Human in the Loop — Watch on YouTube · Listen on Spotify · Listen on Apple Podcasts
Every episode — full show notes, sources, and the archive — lives at podcast.vallyseed.com. The show is produced by VallySeed, where Oscar and I help organizations design, build, and deploy AI systems that create measurable competitive advantage.
