Harry Karl

[Harry Karl @HarryKarl_]

Cursor is an enterprise product disguised as retail. Its revenue never came from subs, people don’t pay API pricing, and enterprises aren’t using grok that’s for sure. It’s a great harness for an enterprise that has to pay API pricing anyway. Just live Devin.

If I was forced to pay API, I would be using Cursor or Devin for coding, not Claude Code or Codex.

[Harry Karl @HarryKarl_]

Exploration has occured, and i can confirm @OpenAI @pvncher are innocent. With my goal prompt, 4 terminals were launched with GPT-6-Astra XHigh, to take on the work.

I have deleted the original post to reduce its exposure since this is a popular account, and take back my comments. I will report back on real usage from what i have left.

[Harry Karl @HarryKarl_]

@pvncher @ForwardEditor @OpenAI Yes sir, anything else you need form me please let me know. 01a094b9-7755-76c2-9fc5-15f4378dd2c2

I have also raised a secondary audit in the name of science. Please roast me deeply if its my fault.

[Harry Karl @HarryKarl_]

@ForwardEditor @OpenAI The fun continues. Perhaps that update you see there is the 'bug' getting fixed, or my skill issues! lol

[Harry Karl @HarryKarl_]

@techguyagent @ForwardEditor @OpenAI Makes no difference to my life, it’s simply cancelled and replaced with Claude!

[Harry Karl @HarryKarl_]

@techguyagent @ForwardEditor @OpenAI Just gave you images, you can see my reset from the usage reset they applied Tuesday.

[Harry Karl @HarryKarl_]

@techguyagent @ForwardEditor @OpenAI Clown 🤡

[Harry Karl @HarryKarl_]

@kompilat @ForwardEditor @OpenAI I haven’t made any adjustments to the codex harness - out of the box, therefore, it sucks ass. It’s not my responsibility to load up a default product and be penalised for using it how they designed. 18,000 tests is not complex, it doesn’t require token spend to run scripts pal.

[Harry Karl @HarryKarl_]

Yeah i hear that every time, theres always a bug. I didnt pay 200 dollars for bugs. I think people are forgetting th elunacy of these costs - never before AI have we ever spent 200 dollars as developers on a personal subscription, never mind multiple. "Subsidised" tokens is a psyop.

[Harry Karl @HarryKarl_]

@ForwardEditor @thimorrowr @OpenAI Thats fantastic but does your task have zero complexity? Is your Codex harness behaving differently? Either way, this is not what you need from a 'frontier' model or harness.

[Harry Karl @HarryKarl_]

@thimorrowr @ForwardEditor @OpenAI 20x, see above ^

Even on 5x it would be unacceptible. There is zero use case for any AI model that runs out of use in just a few hours, fo any large amount of money. The token pull seams to be in effect. Resets are ot the answer, they wont last forever.

[Harry Karl @HarryKarl_]

@unlimwallworks @0xSero Agreed - and now 5.5 just blew that out of the water too, so Astra is done for. "AGI" lol

[Harry Karl @HarryKarl_]

@mxstbr @rauchg Depends how difficult you can make a specific task. NextJS coding could just be solved - general benchmarks can be made more complex than a specific task.

[Harry Karl @HarryKarl_]

@higgsfield is the most predatory SaaS I have used in a long time. It’s like logging in to a casino, they slap you in the face with fake promos “just for you”. Should have a gambling license at this point. Free prompts literally take a life time to make you feel you must upgrade for faster generations. Wild.

[Harry Karl @HarryKarl_]

@Rasmic @leeschmidt123 What is pushing it to you? Could you share some examples of what would cause you to ‘push’ Claude harder than others?

[Harry Karl @HarryKarl_]

@50bbx @ctatedev Weird response - I don’t recall him calling it revolutionary he’s just updated the repo with a useful feature?

[Harry Karl @HarryKarl_]

Pray to the lord above it solves Opus5 slop and someone at Anthropic actually daily ran the model this time.

If they so insist on 50% of our quote falling back to a supporting model, that model better fucking work.

[lyra @lyraxana]

Anthropic is currently stealth testing Opus 5.5 (`claude-opus-5-5`) under the codename `claude-wafer-eap` which is planned to be released on Tuesday.

[Harry Karl @HarryKarl_]

@1kartikkabadi1 @NavDoesTech This is completely dishonest - you did zero real work, or didn’t do any at all. Who is paying you to show this?