Theo-brainUnofficial. Not affiliated with Theo or t3.gg.
Sign inYou
43:15
0:00

Ox Alpha is INSANE

Thursday, Aug 27, 43 min, about 7 min to read

Theo-brain watched all 43 min

Ox Alpha, the anonymous mystery model, turned out to be GLM 5.3 Flash, which Theo says behaves like Opus 4.8 at a tenth to a hundredth of the price.  It is weak on intelligence but strong on behavior: it follows instructions, stays on task and unblocks itself.  Auditing every open pull request in his T3 Code repo cost about 12 cents, while a narrower review on Fable cost over $100. 

Skim itJump to Ox Alpha’s pull-request audit to see the kind of review work Theo gets for cents.

Tap any time to jump the video there. Quotes are YouTube's automatic captions, word for word, so some names come out misspelled.

Key takeaways

The mystery model 'Ox Alpha' turned out to be GLM 5.3 Flash. Theo says it behaves as well as Anthropic's Opus 4.8 at a tenth to a hundredth of the price, and it quickly became one of his most-used models.

In fact, Aux Alpha has quickly become one of my most used models, and I really, really like using it. But, I bet you wouldn't have guessed it's a flash model. GLM 5.3 flash. It behaves better than 5 3, it behaves on par with models like Opus 4.8 at a 10th to a 100th the price.

The cost gap is huge. Auditing every open pull request (proposed code change) in his T3 Code repo cost about 12 cents. Running a similar review on Fable, just for his own PRs, cost him over $100.

It was 12.06 cents.’ Here I was, big spender out here spending 50 cents for useful insights. Nope, it was 12. 12 cents to get a useful insight.I've run this on Fable, and just my own PRs was like over a hundred dollars.

The trade-off: it is not as smart as Fable or Sol and won't solve hard problems as well. But it solves easy problems far cheaper and does real agentic work for you in ways that, he says, other things just won't.

It will not solve hard problems as well as fake or soul will. But it will solve easy problems comically cheaper and it will do real agentic work for you in ways that other things just won't.

Its weak spot is token efficiency: about 47k tokens per task versus 20k for Luna on max. That matters because more tokens means slower work and a context window (the model's working memory) that fills up before the job is done.

3 flash on artificial analysis did around 47k tokens per task on average where Luna on max only did 20k and so only did 17k.If you use more tokens to complete a task your context window gets filled faster and your likelihood of getting to the solution before you max out your context goes down.

Skills in this one.

Cheap-model PR triage

What it is
Point a cheap, well-behaved model at all open pull requests (proposed code changes) and have it prioritize them by how easy they are to merge and how much value they give users.
Why Theo cares
A run cost him about 12 cents for a useful insight. He plans to set up a bot that runs it every 3 hours and pings him when a pull request looks ready to merge, which he expects to change how fast his team ships.
How to try it
  1. Ask your agent: look at all open pull requests in this repo and help me prioritize them by ease of merge and value to users.
  2. Once it works, have it run on a schedule and ping you when a PR looks ready to merge.
So, let's give that a shot. Take a look at all the pull requests that currently have open in this repo. Help me prioritize them based on ease of merge and the value that they provide to our users.
12 cents to get a useful insight. I'm going to set up a bot that runs this every 3 hours and pings me when there's a pull request it thinks is ready to merge. And it's going to change how fast we get shipped.

Advice

Don't default to the highest model or reasoning setting. From running T3 Chat, Theo sees users who always pick 'extra high' burn through their usage without knowing why. Most would be fine with a cheaper model.

People just default to the most expensive, most high-end thing, having no idea what that even means, and then end up burning themselves a bunch. Most of the people who feel that way would be totally fine with a model like Luna.

Write precise prompts. A model that does exactly what you say will do exactly that. Theo asked it to review all PRs in the repo when he meant only his own, and it did just that.

I should have second-guessed what I sent because AUX, aka 53 Flash, does what you told it.

Try GLM 5.3 Flash if you want something very cheap, surprisingly capable, thorough and well-behaved.

If you want something absurdly cheap, surprisingly capable, and incredibly thorough and well-behaved, you should give a model like this a shot.

Hot takes

With Fable, Sol and GLM 5.3 Flash, there is practically no reason to use any other model on his tier list.

His reasoning: His list ranks the unique value a model brings, not raw smarts. 5.3 Flash isn't smarter than K3 or Opus 5, but it is useful in ways the others aren't.

Obviously, 5.3 Flash isn't smarter than K3 or Opus 5. This list is meant to be ranking what unique value does this model bring? Is it actually useful in ways others aren't? And I honestly believe with these three models, with Fable, Soul, and 5.3 Flash, there is practically zero reason to use anything else in this list.

There is no longer any reason to use DeepSeek V4 Flash.

His reasoning: It was compelling because you could run it cheaply yourself, but it has no vision (it can't read images), and on long tasks he found it got distracted and needed nudging.

V4 Flash was really compelling cuz it was cheap when you could run it on your own stuff, but it had no vision. So I have no reason to use that over 5.3 Flash like ever.I also found that on longer running things it would get distracted and lost and need more help and encouragement even the new snapshot…

AI isn't getting more expensive. More intelligence is. At any fixed level of intelligence, costs fall exponentially.

To everybody who's complaining that AI is getting more expensive, I hope you understand that more intelligence is the thing getting more expensive. At a given level of intelligence, the costs go down exponentially.

Gemini 3.1 Pro is extremely intelligent but close to unusable for coding.

His reasoning: Ask it what to change and it answers well. Tell it to fix the code and it reasons through nonsense, loses track, and decides to rebuild your codebase.

If you gave it code and said, ‘What should I change to fix it?’ it would probably do that really well. If you say, ‘Okay, fix it.’ it's going to reason a bunch of nonsense and then say, ‘Okay, I'm going to rebuild your code base.’ because it lost track of what it's doing.

How Theo thought about this

He splits model ability into two tracks: intelligence (how much it knows) and behavior (how well it does what it's told, stays on task, takes new instructions mid-task and unblocks itself). GLM 5.3 Flash is the first model he's found that is weak on the first but great on the second.

…to be blunt, intelligent models are the ones that have the most knowledge. Behavior models are the ones that apply what they're told to the best.What has never happened for me at least before this model is something that is both genuinely stupid but also genuinely awesome to work with.

He admits the anonymous release changed his mind. Had Z.ai only put out a blog post, he'd have dismissed it. Testing it blind as a free mystery model let him judge it on how it actually worked.

Frontier intelligence at flash cost. If they had just put out this blog post without the early access window and without the anonymous aux drop, I would have called on this and probably not even looked. They made a really good call making this model free as an anonymous drop because it made me interested enough to try it and blown…

Sign in to mark what you've watched and keep notes.

Have a question about this one? Ask about this video

Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.