Frederic Koh

[Frederic Koh @KohFrederic]

@Doradora027 @Spideyy_xx @grok sorry bud, religious is 25 cm these days

[Gia @giadotai]

dear openai, can we get some clarity on codex plus limits? even luna max seems to consume more usage than before. thanks.

[Frederic Koh @KohFrederic]

@darekgusto no, i'm saying i'm only willing to be screwed over if the AI lab is clearly frontier

openai doesn't get to screw us over when their models are so much behind

[Frederic Koh @KohFrederic]

You don't have a good understanding of gametheory

OpenAI and Anthropic is dependent on billing enterprise for a huge margin via API. The entire American economy is depedent on the AI story which hinges on those two as of now.

If you're able to serve on-par intelligence at a drastically cheaper price, with transparent weights and reasoning, customization, no bias, uncontrolled by foreign companies, etc...?

Then there'd be no need for AI labs that bill like that.

No soveriegn power would EVER let their lifeline be dependent on a foreign entity, much less willful techbros. Even America doesn't trust its own AI labs, just look at the pentagon vs anthropic.

[Frederic Koh @KohFrederic]

@_jayyy18_13k no usage caps? that's impossible lmao

if that was true i'd subscribe to a yearly one in about two seconds

[Frederic Koh @KohFrederic]

ask yourself this: do you really trust what openai claims is going on the backend?

30k tokens with a "hi" that responds in 2 seconds? deadass try it yourself right now

even if there's some technical reason behind it, my point stands. how can you trust a company known to screw its users over to bill you fairly when you can't see the reasoning, cache, etc.?

no matter what they claim, if you can verify it on your end, don't believe it

[Frederic Koh @KohFrederic]

@homberq @iamsstef lmao i mean openai support is ass, i can tell you that

[Frederic Koh @KohFrederic]

@LuciferMornens @iamsstef how tf did you get your money back lol

[Frederic Koh @KohFrederic]

@Tech_girl uh okay

i dont think people need to know 6.1 sol exists though

it's slower and worse

[Frederic Koh @KohFrederic]

@JamesMalsawm i dont know

i really wish they shipped a frontier model to mog claude

but at this point, i'm not too hopeful of openai

their newest model is likely going to try to compeete with claude's fable 5.5 or opus 6, whenever that releases or at the end of the 28-day challenge

[Frederic Koh @KohFrederic]

@LDT0545 with the fucking 20 tps of sol, i couldn't even burn my usage limits if i tried

[Frederic Koh @KohFrederic]

@hiarun02 switched to claude and don't regret it

[Frederic Koh @KohFrederic]

No. You don't even know how much usage you even have.

Saying "hi" costs 30k tokens for a response on Astra Light reasoning mode.

There's zero transparency. Even if they say the doubled the usage again, can you trust them?

They don't show reasoning nor cache, why would you ever bill API on a second-class model which is not transparent?

Maybe they send the request to a cheaper model or purposely throttle you. I don't believe they've never done that before.

[Frederic Koh @KohFrederic]

@rezoundous yep

[Frederic Koh @KohFrederic]

@hakmgpt 1, 2, and 5 are the only viable options

5 is the guarantee likely eventually, but it's not much. 1&2 are the ones which will cost something

ultrafast may get to 15x speed, only if we get like 150x cost LMAO

[Frederic Koh @KohFrederic]

@danyay same here, literally me yesterday LMAO

[Frederic Koh @KohFrederic]

It doesn't matter how many bug fixes Tibo does.

So what if they say it's more efficient? After all the times they've screwed us over without being transparent, how can you trust them?

The models are slow, there's widespread degradation, reports of models being dumber, etc.

How do we know you're not routing to a different cheaper model?

We can't see the cache, reasoning, or anything. Did you know a simple "hi" message on Astra light reasoning takes 30k tokens? Are you fucking kidding me? Istg, test it out yourself.

Can you even trust that the tokens are billed correctly on the subscription, much less the API? Everything is "proprietary." Where did the "open" of OpenAI go?

The real nail in the coffin for OpenAI is losing trust in a system that's opaque and where you're not in the lead.

Anthropic can get away with it BECAUSE THEY HAVE OPUS 5.5. We can literally FEEL that it's better.

There are honestly only three paths forward for OpenAI:
1. Be outcompeted by Anthropic and open-source models
2. Full transparency of everything.
3. A new fast, cheap, compliant, and intelligent model that is a whole tier above the coming Fable 5.5/Opus 6

I used to be a fan of OpenAI, and I prefer it over Anthropic. But nothing can make up for dishonesty and underperformance.

[Frederic Koh @KohFrederic]

It doesn't matter how many bug fixes Tibo does.

So what if they say it's more efficient? After all the times they've screwed us over without being transparent, how can you trust them?

The models are slow, there's widespread degradation, reports of models being dumber, etc.

How do we know you're not routing to a different cheaper model?

We can't see the cache, reasoning, or anything. Did you know a simple "hi" message on Astra light reasoning takes 30k tokens? Are you fucking kidding me? Istg, test it out yourself.

Can you even trust that the tokens are billed correctly on the subscription, much less the API? Everything is "proprietary." Where did the "open" of OpenAI go?

The real nail in the coffin for OpenAI is losing trust in a system that's opaque and where you're not in the lead.

Anthropic can get away with it BECAUSE THEY HAVE OPUS 5.5. We can literally FEEL that it's better.

There are honestly only three paths forward for OpenAI:
1. Be outcompeted by Anthropic and open-source models
2. Full transparency of everything.
3. A new fast, cheap, compliant, and intelligent model that is a whole tier above the coming Fable 5.5/Opus 6

I used to be a fan of OpenAI, and I prefer it over Anthropic. But nothing can make up for dishonesty and underperformance.

[Tibo @thsottiaux]

Day 12/

We’ve made improvements to Codex & Work that increase the effective usage you get. We achieved this by reusing context across subagents, keeping context cached longer, and improved request routing that allows us to increase cache hits.