Arham

Building a bunch of software and seeing what sticks. Currently Astreex + more. Sharing the work, numbers, and mistakes along the way.

[Arham @arhamsai]

Codex PR reviews are so much better than CodeRabbit and Greptile it's not even close. I wrote a custom skill to babysit a PR and let it run overnight. CodeRabbit and Greptile both gave up after 2-3 fix commits. Codex kept going. 80+ further findings, fix after fix on every commit. It's included in your Codex sub, you just turn it on from Codex web app.

[Arham @arhamsai]

@chrismaconi @cursor_ai The usage limits in Grok bot honestly make it unusable

[Arham @arhamsai]

@SSHCodes Yeah the Gemini 4 timing doesn't add up. Still, whatever this thing is, it's not coming from a small lab.

[Arham @arhamsai]

Starting to think Ox Alpha might just be Gemini 4

[Vidhi @vidhiparmarr]

wtf: Gemini team started a new war

whole team is suddenly hyping Gemini out of nowhere

something big is coming i guess, is the pretraining of Gemini 4 finished?

i think Gemini team is really confident with their internal results for Gemini 4,

it feels right because last month, July 21st to be precise, the team announced that they are really excited from their early progress

Google shocked everyone before

i remember Gemini 2.5 shocked everyone, almost everybody thought it's over for google, but dropped SoTA and mogged every other model

here is my best prediction from this hype

Gemini 4 is actually beating Fable and GPT 5.6 Sol in their internal benchmarks.

plus they must have trained a big model which will outcompete upcoming OpenAI and Anthropic models Astra and Fable 5.1 respectively

looks like we don't have to wait long for it and Gemini is finally making a huge comeback soon,

i am super ready for it.

[Arham @arhamsai]

@teortaxesTex It's not fast at all. Literally giving 17 TPS, but it's most probably not from a Chinese lab.

https://t.co/XKPmqF1uJx

[Arham @arhamsai]

Ox Alpha is probably not a Chinese model.

OpenCode claims zero data retention. A Chinese lab can't offer that. They reported 2.6T tokens in a single day, OpenRouter has 5T total. OpenCode also claims 100T token/day capacity. Only xAI has the infrastructure to offer that kind of capacity for free.

And when you prompt it in Chinese, it reasons in English internally before responding.

This is probably a Grok or Composer model.

[Arham @arhamsai]

@DanDr1s Not sure about the model, but it's probably xAI behind this because only they can offer this much compute. However, I'm sure it's not Grok 4.7.

https://t.co/XKPmqF1uJx

[Arham @arhamsai]

Ox Alpha is probably not a Chinese model.

OpenCode claims zero data retention. A Chinese lab can't offer that. They reported 2.6T tokens in a single day, OpenRouter has 5T total. OpenCode also claims 100T token/day capacity. Only xAI has the infrastructure to offer that kind of capacity for free.

And when you prompt it in Chinese, it reasons in English internally before responding.

This is probably a Grok or Composer model.

[Arham @arhamsai]

@_DMontgomery40 @elshayib_ I don't believe it's Xiaomi because it probably can't offer Zero Data Retention

[Arham @arhamsai]

@elshayib_ Agreed, only xAI has that kind of infrastructure. My guess is it's a new Composer model or a smaller Grok model, not Grok 4.7.

https://t.co/XKPmqF1uJx

[Arham @arhamsai]

Ox Alpha is probably not a Chinese model.

OpenCode claims zero data retention. A Chinese lab can't offer that. They reported 2.6T tokens in a single day, OpenRouter has 5T total. OpenCode also claims 100T token/day capacity. Only xAI has the infrastructure to offer that kind of capacity for free.

And when you prompt it in Chinese, it reasons in English internally before responding.

This is probably a Grok or Composer model.

[Arham @arhamsai]

@Ananth7e I don't think a Chinese lab can provide ZDR

https://t.co/XKPmqF1uJx

[Arham @arhamsai]

Ox Alpha is probably not a Chinese model.

OpenCode claims zero data retention. A Chinese lab can't offer that. They reported 2.6T tokens in a single day, OpenRouter has 5T total. OpenCode also claims 100T token/day capacity. Only xAI has the infrastructure to offer that kind of capacity for free.

And when you prompt it in Chinese, it reasons in English internally before responding.

This is probably a Grok or Composer model.

[Arham @arhamsai]

@ziwenxu_ Most probably xAI. Definitely not Chinese.

https://t.co/XKPmqF1uJx

[Arham @arhamsai]

Ox Alpha is probably not a Chinese model.

OpenCode claims zero data retention. A Chinese lab can't offer that. They reported 2.6T tokens in a single day, OpenRouter has 5T total. OpenCode also claims 100T token/day capacity. Only xAI has the infrastructure to offer that kind of capacity for free.

And when you prompt it in Chinese, it reasons in English internally before responding.

This is probably a Grok or Composer model.

[Arham @arhamsai]

Ox Alpha is probably not a Chinese model.

OpenCode claims zero data retention. A Chinese lab can't offer that. They reported 2.6T tokens in a single day, OpenRouter has 5T total. OpenCode also claims 100T token/day capacity. Only xAI has the infrastructure to offer that kind of capacity for free.

And when you prompt it in Chinese, it reasons in English internally before responding.

This is probably a Grok or Composer model.

[Arham @arhamsai]

Anthropic's ARR came in at $6.5B. Every third party estimate had them at $7.4B+.

Meanwhile Anthropic shipped Fable 5, then OpenAI dropped GPT-5.6 Sol with Codex improvements and resets back to back. Fable is still the best for raw code quality but Sol crushes it on efficiency. Already cheaper, now 50% off on API. Plans give you way more Sol usage than Claude gives you of anything.

Opus 5 is bad. Users are moving to Codex and the revenue is starting to show it.

[Arham @arhamsai]

I have 2 Codex accounts. Both drain in about 2 days now. Haven't even touched the banked reset, just sitting at 0%. Been using Ox Alpha in Pi instead. Cache miss or not, something changed.

[Tibo @thsottiaux]

Update on rate limits in Codex. We do see that for some users the cache hit rate has been worse this week than the stable state the weeks before. This could explain that usage is draining somewhat faster for those users as hitting the cache consistently is an important component of being efficient.

We are investigating and will have an update tomorrow.

[Arham @arhamsai]

@bridgemindai Same, didn't get the last reset either

[Arham @arhamsai]

The best AI builders I've seen all have the same background. Years of writing docs, filing detailed bug reports, explaining requirements to offshore teams. Turns out communicating precisely was always the skill that mattered. AI just made everything else optional.

[Arham @arhamsai]

@maxktz @linear Gonna give this a try

[Arham @arhamsai]

Ox Alpha just carried me through 3 PRs, 95 files, and ~3200 lines of code. Swapped an entire AI provider, rebuilt how background jobs dispatch, fixed a search system that was matching the wrong results. This model actually ships.

[Arham @arhamsai]

Ox Alpha in OpenCode was rough. Slow, kept stopping mid-task, I had to keep telling it to continue. Switched to Pi a few hours ago and it hasn't stopped once. Faster too. Same model, completely different experience depending on where you run it.

[Arham @arhamsai]

@theo It works much better in Pi compared to OpenCode.

[Arham @arhamsai]

@rakesh_rry @dannypostma @aarondfrancis Looks cool. Any way to try it out?