Ox Alpha – Use This While It's FREE (Stealth Model)
BBetter Stack
Computing/SoftwareInternet Technology
Transcript
00:00:00right now there's a top tier model that you can use completely free and nobody knows what it is
00:00:04it's called oxalpha and it's on open router open code klein and news portal for free it's got a
00:00:081 million token context window it takes text images and videos and it gives some seriously
00:00:13impressive results and even though nobody will say what it is we do have some pretty good clues
00:00:22so this all started when open code tweeted oxalpha stealth model free for the next week
00:00:261 million context multimodal zero data retention and they have capacity for a hundred trillion
00:00:31tokens per day this was followed up by open router making the same announcement and news research
00:00:36saying they have capacity for one quadrillion tokens per day like damn this is a seriously big
00:00:40stealth launch now before we jump in and play detective it's worth asking is it even worth
00:00:44doing that is this model actually good and the answer is mixed klein tweeted this benchmark showing
00:00:49it beating fable and soul by quite a bit on deep sw bench but this is misleading in my opinion and
00:00:54anyone telling you this is better than fable is either click baiting or overhyping it and i'm
00:00:58not here to do that so subscribe if you appreciate that the source for this was actually ben's tweet
00:01:03where he ran it through 10 of the tasks on deep swe and yes it did beat gpt 5.6 on those but on a
00:01:08more complete run it's looking closer to 63 which puts it near grok deep seek and gemini or fable
00:01:14and 5.6 soul on a medium effort level so keep that in mind for your comparisons something else
00:01:18interesting though is apparently averaged 47 000 output tokens which groups it closer to the closed
00:01:23models but i must say i don't think it is one there is just so much speculation on twitter out there at
00:01:28the moment so instead of comparing all of these competing benchmarks let's just jump in and try it
00:01:32ourselves so i ran the usual tests that i run on all of these models when they're released the first
00:01:36one is for a personal finance management dashboard with a load of features but i don't tell it the
00:01:40stack to use i just want a full stack application because i'm pretty curious at what tech stack these
00:01:44actually choose you can see here i used aux alpha used 84 000 tokens it's actually ran for 45 minutes as
00:01:50well but the good news is this one is completely free i also compared it against fable medium soul
00:01:55medium deep seek glm 5.3 gemini 3.7 flash and kimmy k3 because i'm pretty certain this is an open model
00:02:01so i want to compare it against them but i'm also curious how it performs against those frontier ones
00:02:05this was the result i got back from aux alpha in that single prompt and i've got to say i'm pretty
00:02:09impressed this is an incredibly usable app i like the ui it's pretty sort of shadsy encoded all of these
00:02:15pages work i've checked that the features work and overall as i said it's a usable app i could use this
00:02:20on my own if i was making a piece of personal software i would not complain if it looked like
00:02:24this when it comes to the actual code it chose a pretty reasonable stack as well it used next js which
00:02:28i can't really complain with and it also used drizzle and better sqlite which is a very good move it used
00:02:33a real database which a few of these models don't so overall i'd say this has done a really good job on
00:02:37this personal finance dashboard app it chose a good stack and the app is usable next up we have what
00:02:42deep seat v4 pro gave me and i've got to say when i first saw this i thought aux alpha might be one of
00:02:46the deep seat models because these uis are incredibly similar and again though this is a
00:02:50very usable ui i couldn't complain about this it looks very shad cn app and all of these accounts
00:02:55transactions and goals tabs do work except they're not actually tabs they're just going to a different
00:02:59point on the page so that is a little bit different there from aux alpha but the uis are
00:03:03incredibly similar and so is the code when we look at the stack that it chose it also chose next
00:03:08js and better sqlite the only difference is it didn't use drizzle which just went its own way by
00:03:12using the database itself i would have preferred to have that drizzle layer deep seek actually took 15
00:03:16minutes on this task and cost 19 cents and while there's a lot of similarities to ox alpha i don't
00:03:21think it is ox alpha and you'll find out why in a bit continuing with the open models this is what
00:03:25kimmy k3 gave me which a lot of people think is the best open model of the moment and this one
00:03:29actually took 38 minutes and cost me a dollar and 92 cents so we've had quite a step up in price on
00:03:34this one again it looks pretty similar to what we have with ox alpha ox alpha is definitely a chinese
00:03:39open model or at least it is as the base as they're all using this sort of same ui design you can see
00:03:44it all works it has the nice dashboard we've got an accounts page transactions and goals this one is a
00:03:48bit more similar to ox alpha where it's actually done separate pages for those but again everything
00:03:52looks pretty standard and it's usable where kimmy differs quite a lot though is in the code it just
00:03:56went with a plain react front end it doesn't have any routing or anything like that and for the back
00:04:00end it chose an express server and it used node sqlite for the database so it does have a database
00:04:05but to be honest with you for one dollar and 92 cents i prefer what ox alpha and deep seed gave me
00:04:10which was a full next.js app that i could scale in the future now i'm going to take a quick detour
00:04:14from the open models and focus on the closed ones this is what fable medium gave me obviously being
00:04:19on that medium level it doesn't want to think for long so it only took three minutes and 42 seconds
00:04:23but it still cost me two dollars and sixty cents which was the most expensive out of all of these
00:04:27and the results are pretty underwhelming it's just a single page it does what it says it does i can send
00:04:32money i can add funds all of that is working but this is just a plain react app with an express
00:04:36backend and it doesn't even use a real database it stores all of this in javascript so i refresh the
00:04:41page everything goes away i think the takeaway from this is if you're not switching to some of
00:04:45these open models to save yourself some money on that higher thinking you are definitely missing out
00:04:49it's a pretty similar story for gpt 5.6 soul on a medium effort level it's a fairly nice ui this is
00:04:54very gpt coded it loves all of these monospace eyebrows and also these side borders but i can't
00:04:58really complain at how the ui looks the features though they're not actually working we can't go to
00:05:03the accounts tab transactions investment or goals hasn't implemented them you can send and add funds
00:05:08but it's actually built this on a next js app but then hasn't used a real database again it's just
00:05:12storing all of this in javascript and that's probably because it just spent four minutes on this and it
00:05:16cost me 33 cents the last closed model that i tried was gemini 3.7 flash and this was on a high
00:05:21level since it matches in the benchmarks and you can see this is a very similar ui to what the open
00:05:26models have been given us i sort of call this the last generation of ai ui it always has this similar
00:05:31look and fable and open ai seem to be moving away from this everything in this is working though and
00:05:36all of these tabs do actually work and there is something on them so bonus points for gemini 3.7
00:05:41flash there this actually took me around 10 minutes and cost 50 cents when it comes to the stack that it
00:05:46chose this is actually a next js application but it didn't use a real database so it's pretty similar
00:05:50to gpt 5.6 onto our last comparison then and this one is glm 5.3 and i left this one to last for a
00:05:56reason you see the ui is pretty nice it went in a different style from the other ones still has sort
00:06:00of this green that all of them seem to be using this one actually went with a light mode and when it
00:06:04comes to the features it's pretty similar to all of the other ones that we had and all of these are
00:06:08working this one actually took around 50 minutes to generate and turns out it cost me five dollars so
00:06:13this is actually the most expensive one but the interesting part is in the stack that it chose
00:06:17it chose a next js application it used better sqlite and it also used drizzle so it's the exact same
00:06:23setup that aux alpha used when i saw that i decided to ask claude to compare all of these code bases
00:06:28and see if there's any similarities and see what it thinks the stealth model is this is a summary of
00:06:32claude's findings and as you can see it thinks the stealth model is a glm one looking at the tests
00:06:36that claude did one of them was to compare structural patterns so this is things like how components
00:06:40are assembled and it found the glm 5.3 has 22.4 in common with the stealth model which is higher than
00:06:46all of the others the only one that comes close is deep seek with 19 it actually notes this could be a
00:06:50false positive since drizzle was used in both the aux alpha and glm 5.3 and they're the only ones to
00:06:55use it if you actually remove drizzle deep seek actually has the most in common with our stealth
00:07:00model so could this actually be a deep seek model well the answer is probably not because next
00:07:04it looked at some rare markers these are non-obvious choices in each code base and it found that stealth
00:07:08and glm 5.3 shared a lot of them they both used a pass amount to sense function and the function was
00:07:14practically identical they both used drizzle orm they also both rewrote the readme none of the other
00:07:18projects did this these were the only ones to rewrite the readme and the readme itself has an identical
00:07:23structure so it's pretty likely from all of this analysis that this is a glm model or at least it used glm as a base
00:07:29you don't just have to take my word for this a lot of other people have looked into this and they found
00:07:33that the tokenizer is practically identical to glm's it gives the same error messages as glm models and
00:07:38perhaps the strongest clue is the video encoder in this model uses the exact same tokens as the glm
00:07:43vision model i'll leave this page link below that has a list of the other similarities but i think
00:07:47it's pretty clear this is a glm model this could be glm 5.5 which is now multimodal or perhaps it's a
00:07:52post-trained version of glm 5.3 or even more exciting this could be a flash version it doesn't even have
00:07:57to be zai either since glm is open this could be another company that's taken glm and used it as a
00:08:02base overall though it's just a very competent model and it has a price that is impossible to beat
00:08:07and that's going to last till the 27th so get in there while you can just know that prompts and
00:08:11completions are retained by the provider but apparently they're not used for training what do
00:08:15you think this model is let me know in the comments down below while you're there subscribe
00:08:19as always see you in the next one
Community Posts
No posts yet. Be the first to write about this video!
Write about this video