Harry Y

Content Creator | Ex-Web3 Strategic Advisor & Market Maker | Bridging Trading Tech with Institutional Narratives

[Harry Y @HarryY4ng]

@mylifcc 之前在Chat/Cowork里用过旧版Project的号是今天才有的，我的旧号就是这样。

之前Cowork很多的Project今天才能升级，Code的Projects今天才出现。另一个新号是建号之后就有了

[Harry Y @HarryY4ng]

@Khazix0918 个人会觉得 注册/首登地区-注册方式-付款方式，比单纯看IP更重要一些。我每年回国大概2-3个月，期间直接用商业vpn的专用IP，其余什么设置都不改，时区系统自动，目前两个20x的号都没出过问题

[Harry Y @HarryY4ng]

@chasen_liao a社的50%限制，我两个Max加起来相当于100%的Fable用量🤣，这是其中一个号，用了两天的状态，一个项目用Fable 5.1 Max在做架构，还有另一个小项目用的是Opus5 xHigh

[Harry Y @HarryY4ng]

@MaxForAI 全体被迫挂机等GPT6🤣

[Harry Y @HarryY4ng]

@buildwithrajath all waiting for Astra, forced, lol🤣

[Harry Y @HarryY4ng]

Claude, Codex, Grok
All down
What a Day🤣

[Harry Y @HarryY4ng]

On cost, the complaints check out.

I ran it at max intensity for about an hour.

That wiped my 5-hour limit and ~20% of weekly credit, which is about 40% of my Fable allowance.

[Harry Y @HarryY4ng]

Firsthand on Fable 5.1.

Some of the controversy is real.

Setup: Max 20x, Claude Code in VS Code, working on a quant system.

[Harry Y @HarryY4ng]

I had Fable 5.1 review work that Opus 5 already did last week -- including multiple rounds of adversarial review.

Credit where it’s due: the model is still top-tier. It caught issues Opus 5 missed.

[Harry Y @HarryY4ng]

@OpenAI just announced it will stop supplying GPT models to @cursor_ai after @spacexai acquired the company. Direct access is set to end on November 12.

Judging from the replies, a lot of people aren’t buying the official explanation.

@thsottiaux jumped in to manage the fallout, not with another reset this time, but with workarounds: bring your own API key, keep using OpenAI’s IDE extensions, and so on. The reset habit is still there in the background. It just wasn’t the tool they reached for tonight.

@elonmusk and the Cursor side responded fast. Elon basically said he couldn’t care less, and called it “Scam Altman.” They have some basis for that confidence: Grok has been getting better, and as SpaceX leans harder into AI, the compute and talent advantage becomes more obvious.

What surprised me more was @AnthropicAI . It wasn’t a formal company blog, but two senior people moved quickly: Cursor has been a trusted partner since Sonnet 3.5, and they will keep increasing compute for Claude inside Cursor. That lines up with SpaceX already supplying compute to Anthropic.

This again fits a judgment I’ve had for a while: Anthropic might be the next Apple, a stubborn, brilliant CEO, unconventional product taste, and a following that borders on religious. A lot of people hate that comparison. Fine. Go back and look at Steve Jobs’ Apple. It was far more eccentric than the polished version people remember now.

I canceled Codex Pro last week and got a refund, keeping only Plus because I still need a few Codex CLI-based workflows. I used the refund to buy another Claude Max on my wife’s account. Codex did help me. It also gave me a few very bad weeks. Lately the reset marketing has started to feel like the product. Not every problem can be fixed by pressing that button. Right now their biggest bet is Astra. They’re selling its capabilities hard. Let’s wait and see.

[Harry Y @HarryY4ng]

Starting around Thursday, Codex became increasingly unreliable and started burning through usage at an abnormal rate.

I had a relatively simple project scheduled to launch on Friday: pull data from Google Search Console via the Google Cloud API to analyze performance across different language sites, plus collect and analyze trending events from media RSS feeds (most of the code could be reused from a previous project).

Codex completely failed to deliver.

I used Plan mode, revised the SPEC and completion path multiple times on GPT-5.6 Ultra, then set a Goal and went to sleep. By morning it had completely exhausted my 5x subscription from 100%, without even finishing a single interface.

To meet the deadline I upgraded to 20x and continued. Same result: it kept making the exact same mistakes over and over.

Today I finally gave up. I canceled the subscription and requested a refund. I don’t care if @OpenAI blacklists the account. I just want to share an honest user experience.

I also use @claudeai and @grok regularly, so I know this level of performance is not normal and not acceptable.

OpenAI just hit 20 million users with Codex, congratulations on the milestone.

I’ll be one of the people stepping away, at least for the short term. Good luck.

[Harry Y @HarryY4ng]

Codex is so fucked up today😅

[Harry Y @HarryY4ng]

Many people are posting this chart saying @claudeai Code / @AnthropicAI has “hit the roof.”

Here’s my take👇

[Harry Y @HarryY4ng]

@Lu00ji rug?

[Harry Y @HarryY4ng]

I genuinely don’t know which one wins.

But I’m fairly confident the ultimate winners will be users, either through better models or lower prices (or both).

Curious what others think.

[Harry Y @HarryY4ng]

Many people are posting this chart saying @claudeai Code / @AnthropicAI has “hit the roof.”

Here’s my take👇

[Harry Y @HarryY4ng]

This leaves two plausible paths:

1. Diminishing returns: the first-mover lead slowly flattens and gets overtaken.
2. Compounding returns: technological + data + financial advantages keep widening the gap.