Aziz

Staff engineer by day, solo builder by night. Shipping with AI agents and posting what works

📈

[Aziz @angaziz]

Beacon was initially exclusive to claude code, but v0.4.0 is the release where I set it free!

Also added a unified settings window on the hub for better experience and bug fixes

Setup is easy if you want to try: web-flash the device and drop the hub app to your mac

GH repo link below 👇

[Aziz @angaziz]

neutrality is easy to promise as a startup. harder after a payments company owns the router

the test isn't the blog post. it's whether routing still looks the same when a stripe-favored model and a cheaper rival are both "good enough"

[Aziz @angaziz]

stripe agreed to acquire openrouter

openrouter's post: 10+ trillion tokens a day, 400+ models, 10 million developers and companies. stripe's: 80+ providers. name, product, and "route for the user" stay the same

[Aziz @angaziz]

people keep arguing which model wins. the acquisition treats that as the wrong layer

if no model is best at every task, the valuable system is the one that picks, meters, and fails over. that's why a payments company wants a router

[Aziz @angaziz]

@ns123abc These kind of guys are exactly the reason ai labs making weird rules that's unbeneficial for most

[Aziz @angaziz]

deepseek shipped vision on v4 flash this week: deepseek-v4-flash-vision-exp

jpeg, png, gif, webp. after resize, each image is capped at 384 tokens

vision was never optional on a coding model if you wanted it to be great. anthropic has been reading screenshots for a long time. the lag was getting that loop right on a flash-class api, not discovering that pixels matter

[Aziz @angaziz]

I still don't understand the stories of agents running 24hrs+, gpt family kept asking questions mid-session while anthropic can be quite lazy

[Aziz @angaziz]

@kimmonismus not alr good, but good enough. And it's easier to justify less capable models with x times cost savings, but not vice versa

[Aziz @angaziz]

@Kavya_learns 1st party subscription gives more generous usage for same amount of $ for the preferred model family

[Aziz @angaziz]

deepseek's api now charges by time of day: off-peak tokens cost 50% less starting aug 16.

Token prices just got time zones, and that quietly reinvents batch computing. The moment inference has a cheap window, every non-interactive agent job (evals, refactors, backfills) becomes a scheduling problem. Cron is back as a cost optimization.

[Aziz @angaziz]

@zapactugta6 man, we talking physical media like mad while all of our photos and videos are all digital smh

[Aziz @angaziz]

@kunchenguid Arguably better than opus but definitely not better than Fable, specifically on planning and ideation

[Aziz @angaziz]

@de3dsoul When gaming was fun

[Aziz @angaziz]

@kimmonismus It's an excellent model, true. But it gave me wrong answer multiple times during planning

[Aziz @angaziz]

@Samaytwt "anti social social club" sticker on the wall

[Aziz @angaziz]

@michael_kove Promising for general users, no i don't think 300 is worth it for a bot where you can't even pinpoint which grok model to use

[Aziz @angaziz]

@g_vered @lolawajs Less probable, esp with bots being easier to setup

Although on the 2nd thought, not unlikely both party uses bot to reply lol

[Aziz @angaziz]

@neo_subhamoy usually banks lol

[Aziz @angaziz]

@drsuffy Wild, all with one eye

[Aziz @angaziz]

@XFreeze mmm, not really. I have more flexibility with hermes

But grok is very promising for a product that just launched, might revisit in a month or so

[Aziz @angaziz]

@catalinmpit Same, i tried to brainstorm and plan but it gave me straight wrong answers