Ox Alpha vs Opus 5 i asked both models to build rocket engine using three.js ox-alpha is not even close to opus 5 performance and it took equal or more time for the output prompt used for both: "Create a 3d SpaceX raptor engine using three.js, use white theme... background so the main object is clearly visible" both are one-shot outputs with no additional changes what do you think about Ox Alpha?show more

I think you are a hater, the model is amazing and antrhopic is shitting their pants now for real. I find this model insanely good

no doubt it’s a good model but it’s no way as good as Claude fable 5 or Opus 5

I think it is, in my 3d testing is really good

Who cares if it's not like fable, its free! And once free is gone, it will be cheap. We don't want fable, we don't have endless money to spend on fable and these extra expensive models. We want a great balance. Grok is a great balance, and Ox is free for now.

It's speculated to be GLM 5.3 Flash, and if a Flash model can come this close to frontier then it's really impressive and frontier labs indeed should be worried Evidence: 1) EXACT same tokenizer as GLM family since GLM 4.5 2) Upstream error message: identical with GLM 5.3's error string, this is infra level setup, rules out third party fine tuning possibility like Cursor's composer 3

also GLM 5 was codenamed Pony Aplha, very similar to this Ox Aplha

It's said that Chinese AI influencers where shocked when they knew the size of the model. I am guessing WAY smaller than GLM 5.3... Possibly much smaller than DeepSeek V4 Flash?

ox alpha is really really dummy, tested it with some evaisl and its really dumb, must be a chinese model.

that might really be a chinese model looking at the way it responses

That IS a SpaceX raptor engine from Ox. It’s just from the future.

😭😂

“I did one prompt to two models that each represent different sides of the most impressive current technology known to man. This ONE result says EVERYTHING you need to know” this is such a whack take it is silly

As many users would agree, ox alpha is an outstanding model and rivals fable 5, not sure if you’re Anthropic Pilled or ignorant trying to ignore the fact that the chinese market is taking openai and anthropic from behind

But if we consider the price 💀

Ox is free 🥀

This build is a good eval for model performance

opus all the way man

no one beats Opus 5 in terms of performance quality

Opus 5 still ahead, no surprise

I won't be paying particularly for Ox Alpha. It does not live up to the hype.

And you forgot to mention that Claude burns through your weekly quota with just one prompt like that, while ox-alpha is completely unlimited and free (at least for now).

yeah that tracks. ox alpha still feels flash-tier. opus 5 just builds the thing

Maybe because Ox Alpha does not have subagents and the ultracode option? What harness was used?

See how good grok does. With spacex eng data in training set result could be interesting

I did with Grok 4.6 and it's pretty good

That's my feel too. Very good model but doesn't feel "frontier" on complex tasks.

haven’t used ox alpha before but with my little experience with opus i’m not surprised it clears

I'm really liking this model so far. It found vulnerabilities in the code on the first try that Fable 5, Opus 5, Kimi K3, and Qwen 3.8 couldn't find twice. And, unlike others, it worked an order of magnitude faster (at least 5 times faster than Chinese ones).

On my test ox perform better

honestly, i’d need to see a lot more from ox alpha before taking it seriously for agentic coding,this result isn’t very convincing

most importantly we want to see who is behind that model

just use @mattpocockuk skills and try again

Hi. How can I use this ai locally in PC improve my work? Can you tell the process d'installations please?

Cursor 3 trained on GLM

Slop teat

What harness?

You're right, Ox Alpha is not even close to opus 5. I've been using Ox Alpha for a while and it's shit. I got more pleasure from using Gemini 3.7 Flash.

show the output bc i bet opus' code is just longer not better. sometimes concise wins

I find the same lol its not that much good in these task. I have been trying to generate a decent looking html file and its keep fucking it up.

Ox alpha is so obviously a Google model

What harness did you use for both? I find massive gaps in performance depending on the harness

Ox Alpha . from China Model.

Let us wait and see what it is under the hood.

Qwen3.8-27B. Source:

one render mostly tests three.js priors, not general coding performance across models