Timothy Cardoza

a guy

[Timothy Cardoza @TCCardoza]

@theo Least it's your left arm... Hopeful only your left?

[Timothy Cardoza @TCCardoza]

@theo Ohno

[Timothy Cardoza @TCCardoza]

@yacineMTB What IS a harness harness tho?

[Timothy Cardoza @TCCardoza]

@theo @yacineMTB Not quite a harness still (although apparently working that direction)

[Timothy Cardoza @TCCardoza]

Update, I like this theory best so far:
https://t.co/sW51y4yKiH

[Gael Breton @GaelBreton]

I’ll throw my hot take in the hat.

I think Ox Alpha is Composer 3.

Based on GLM (the tokenizer clues), served by xAI (100T tokens/day)

[Timothy Cardoza @TCCardoza]

My guess is this model is Grok. It's just a hunch but US Eastern, about a week after GLM 5.3 just came out. A week before Grok 4.7 is expected. Claiming 100T tokens per day of compute (both not a huge model and lots of compute power). Everything points to Grok 4.7

GLM 5.x and Grok 4.5/4.6 models always felt similar anyway. Also image and video input matches Grok more than GLM (although this could obviously change with a new model)

[Niklas @niklasdev]

OX Alpha is hosted in America!! 🇺🇸

I sent 49 identical one-token requests from 7 Cloudflare regions, rotated the order to control for load and matched every response to OpenRouter’s per-generation server timing.

This isolated the geographic network floor:
ATL: 57 ms
DFW: 76 ms
Europe: 125–132 ms
Asia: 202–340 ms
Australia: 279 ms

A robust 6-of-7 fit converged at 43°N, 73°W, with the plausible band covering the US East Coast into southeastern Canada.

Best interpretation: OX Alpha’s serving path is anchored in eastern North America, most likely US East.

The only potential remaining caveat: this could be @OpenRouter backend handoff rather than the physical GPU rack.

[Timothy Cardoza @TCCardoza]

@GaelBreton That's actually the most likely guess I've heard so far. I kept thinking it sounded like Grok 4.7 but then it would do small model things and would fingerprint as GLM

[Timothy Cardoza @TCCardoza]

@vamsibatchuk At this point, maybe a new flash model every 3 weeks is all we need 😂

[Timothy Cardoza @TCCardoza]

What the heck is Grok 4.6 low doing way over there??

[Timothy Cardoza @TCCardoza]

@thegenioo Sad for Terra. Over API, Sol is just a little bit better of an option for most things now

[Timothy Cardoza @TCCardoza]

@rezoundous "all we need is the world" 😂

But also yeah

[Timothy Cardoza @TCCardoza]

@maria_rcks Tibo claims it is not. Idk tho, lots of people track their usage

[Timothy Cardoza @TCCardoza]

@FelipeFr1702 Every once in a while it goes a mile in the wrong direction. You might stop it and ask another thread to evaluate that thread (you can right click on a thread to get it's thread id and hand it to another thread and they can actually read other threads)

[Timothy Cardoza @TCCardoza]

I spend way too much time studying various ai models (LLMs ASR models TTS models etc.) and harnesses and their designs.

Is there anything y'all wanna know or want me to research for you? Any theories you want me to look into or hunches you want verified?

[Timothy Cardoza @TCCardoza]

@s_batzoglou Sounds like the kind of hype a ghost model would get. Ig we'll see

[Timothy Cardoza @TCCardoza]

@mdam10x Meee

Let's connect! What are you building

[Timothy Cardoza @TCCardoza]

Not sure how I feel about the cheaper Sol. 20% cheaper over API is awesome. But poor poor Terra...

Apparently, unless there are additional deals, use Sol low or medium over Terra if you're over API.

If you're on sub:
Luna high < Terra High < Sol medium < Sol high < Sol xhigh (no max or ultra, it's illegal 😂)

[Timothy Cardoza @TCCardoza]

@Im_IrushiK Not quite true. The DeepSWE test done by Ben was JUST THE FIRST 10 TASKS. Not a valid full result. Enough to give you an idea that it is at least frontier on some problems

[Timothy Cardoza @TCCardoza]

@zebassembly Can't say it writes the best code... But it is very usable and very fast. Since I get it for free I have Sol use it as a sub-agents through the terrible CLI. Very messy, kinda works. Sol tears Gemini's code apart but also the end result is better than just Sol

[Timothy Cardoza @TCCardoza]

@Ananth7e Well, they already reduced their margins massively for Codex users so... Fair enough ig

Codex sub is still like 700% cheaper than api