Our Claude Code Is Different — The Truth About Three Ways In

90% of the creators online teaching Claude Code aren’t teaching the real Claude Code at all. If you genuinely teach Claude Code, there’s a good chance the video won’t go out — mine get throttled on plenty of platforms the moment they post.

People who’ve watched my last few episodes keep DMing me the same question — “Pinpin, why is the Claude Code I set up completely different from yours?” It doesn’t feel the same to use, and it doesn’t even feel as smart.

There are actually two reasons.

The first is the difference between Skills and the harness — after you install Claude Code, you first have to configure a set of operating rules for it. Like a kid just out of school, you have to teach them the principles of the work and how to communicate with you. I’ll do a dedicated episode on that later.

The second, and today’s topic, is how the way you connect to Claude Code changes what you get out of it.

Open any video platform right now, search “Claude Code,” and watch a creator spend half an hour or an hour walking you through the install — most of what they’re setting up isn’t Anthropic’s official model anymore. Anthropic opened up third-party model access a while back, so a lot of what’s being taught now runs on domestic models.

Not that domestic models are bad — MiniMax, Zhipu, Kimi are all solid models, very usable for Chinese orchestration and research. But we have to admit there’s still a gap. Only once you see the gap can you learn from it and work on catching up.

Even with a genuinely official, direct API connection, the difference from a subscription plan is quite noticeable. Today I’ll pry this open and explain it in plain language.

Three ways in

There are three ways into Claude Code.

The first is the official subscription — monthly, at $20 / $100 / $200. If you use it heavily more than three hours a day, the monthly plan is the better choice.

The second is the official API — you pay per token (word unit) for what you send and what it sends back. This fits low-frequency use, or building a product on top of it.

The third is relay APIs and third-party proxy account pools. Plenty of these — setting aside how shady they are — will quote you an absurdly steep discount. But once you actually use them, never mind that their “IQ” is lower than the official model, their real cost may be higher than official too.

Where’s the difference? I’ll give you the three that matter most.

Difference one — it can’t do the work

Claude Code is called Code not because it can only write code, but because it gets a lot of work done for you through code and scripts. That process is called tool use — it reaches for a lot of tools: reading files for you, running commands, committing code changes into git. That’s it working with its hands.

On the official API, chat and tool use run through the same pipe. Once a relay puts another hop on that pipe, all kinds of problems show up.

You tell it to read the project code, it chats with you perfectly well, and the moment it actually opens a file it throws an error — most commonly a 400. You ask it to fix the code and re-run the tests, the first tool use is fine, but by the time it runs the tests it’s forgotten why it was doing this in the first place. Multi-step tool use produces even more problems.

Very often the first message is fine and the second one dies outright. Because Anthropic puts an encrypted seal on the model’s thinking process — one relay hop and that encrypted part is lost. It looks like it’s doing work for you, but it isn’t delivering 100% of its capability. The gap is real.

Difference two — no memory

The second one is about prompt caching.

What makes Claude Code good is its context — how much of your conversation it can hold. It’s 1M context. Throw a project of a few hundred thousand lines at it, and once it’s read it through, it doesn’t have to read it again to do the next thing. The technical term for that capability is prompt caching.

The official cache is free for one hour. Have it read a 200,000-character novel, and within that hour, as long as the prompt is the same, re-reading costs nothing.

But once you’re on a relay — they now inject a random string, which effectively invalidates the cache. Meaning it re-reads that 200,000-character novel every single time. And the relay may charge you many times over for it.

It looks like a 70%-off deal, and in real use it may cost you more.

Difference three — the model gets swapped

If you’re on a relay with no brand or a shaky reputation, you think you’re connected to Opus 4.7, and behind it they may be using Haiku, or Sonnet, or even some domestic model as a stand-in. On a simple search task, they may just swap the model out — and you have no idea on the front end. That’s a business-integrity problem.

Then there’s data security. Never mind poisoning. On a lot of relays, everything you send is fully visible to them in one shot. They could steal your code, steal your documents — that’s actually the lesser concern. The bigger one is that they’ll take your information to train on and to sell. You have no idea whether a good idea of yours, or some confidential file, gets exposed along the way.

The poisoning I mentioned means this: they can slip fake tools into your instructions — it looks like it’s executing the task you asked for, while quietly doing something else.

People in this space are too clever by half.

That said, it’s not that no relay can be trusted; some are usable. I won’t recommend any here, because I don’t use relays much myself — usually only for video models or image models, where there’s not much privacy to expose.

April 4 — the hard cutoff

On April 4 this year, Anthropic cut off all third-party client calls — tools like OpenClaw and Hermes can no longer be driven from a subscription. So prices all went up too.

The official subscription is the best way

So if you’re going to use Claude Code, the best way is still an official subscription account.

The $200 plan — unless you’re pulling all-nighters developing, you actually can’t max it out.

Do the math at official API rates. Anthropic once gave me some credits, $200 of extra usage on the $200 plan. I turned on Opus 4.7 + fast mode, and burned through $200 in a bit over 20 minutes. Even without fast mode, a little over an hour would take out the whole $200.

So — don’t mind the hassle, don’t go looking for the so-called easy route. Let’s take the official subscription, which I currently think is the best tool available, and actually use it.

Source: EP0017_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the 8:07 original recording

[00:00] 90% of the creators online teaching Claude Code aren’t teaching real Claude Code at all. If you actually teach Claude Code, there’s a good chance the video never gets out — like my videos on a lot of platforms, they go up and immediately get throttled.

[00:10] Friends who’ve watched my last few episodes keep DMing me the same question — “Pinpin, why is the Claude Code I’ve set up myself completely different from yours? The feel of using it, right, and it doesn’t even seem as smart.”

[00:22] There are really two reasons for that. The first is about the difference between Skills and Harness — after we install Claude Code, we first have to configure a set of operating rules for it. Like a kid who’s just entered the working world, you have to teach them some workplace principles and some basic ways of communicating with you. I’ll do a whole episode on that later.

[00:44] The second one, which is what I’m covering today, is how the way you connect to Claude Code affects using it.

[00:51] Open any video platform right now and search Claude Code, and the vast majority of creators — half an hour, an hour, whatever — the Claude Code they walk you through installing step by step mostly isn’t Anthropic’s own model anymore. Because it opened up third-party models a while back, so a lot of what people are teaching now is domestic Chinese models.

[01:08] Not that domestic models are bad, right. Look at MiniMax, Zhipu, Kimi — all solid models. For some Chinese-language orchestration and some research work, they’re still very usable.

[01:21] But we have to admit there’s still a gap. And only once we see the gap can we learn from it, take stock of it, and find a way to catch up.

[01:32] Beyond domestic models, these creators will also teach you to configure some API, or something claimed to be an official direct connection. And even if it really is an official API direct connection, the difference between that and the subscription plan is pretty noticeable. So today I’m going to crack this open and try to explain it clearly, in plain language.

Three Ways In

[01:55] First off, there are three ways to connect to Claude Code.

[01:57] The first is the official subscription — you pay a monthly fee, and there are three tiers: $20, $100, and $200. If you’re basically using it heavily for three or more hours a day, the monthly plan is the better way to go.

[02:09] The second is the official API, where you pay by the tokens — the input you send and the reply it sends back. That’s the connection you use for low-frequency usage, or when you’ve built a product that needs to call it.

[02:25] The third is relay APIs and third-party proxy account pools. A lot of these — never mind whether it’s gray or black market — a lot of them will tell you they’re offering a huge discount. But once you actually use it, forget for a moment that its IQ is lower than the official model, the real cost may end up higher than official.

[02:45] So where exactly is the difference?

Difference One — It Can’t Get Work Done

[02:47] The reason Claude Code is called Code isn’t that it can only write code — it’s that it can use code and scripts to get a lot of work done for you. That process is called tool use, meaning it uses a lot of tools.

[03:00] It can read files for you, run commands, and put code changes into git — those are all things it does hands-on, including running scripts.

[03:11] With the official API, chat and tool use go through the same pipe. Once a relay re-routes that pipe, all kinds of problems show up.

[03:20] Say you tell it to read the project code. It might be chatting with you just fine, saying it’s going to go read — but when it actually opens the file it throws an error. The most common one is a 400, right, error 400.

[03:29] And when you say, now that you’ve changed the code, run the tests again for me, the first tool use might be fine, but suddenly when it runs the tests it forgets why it was doing this in the first place. And then when the tool use has to happen across multiple steps, a lot of problems come up.

[03:45] It’s often fine on the first turn, then dead on the second. Because in the model’s thinking process, Anthropic actually attaches an encrypted seal — and when a relay passes it along, that encrypted part gets dropped. You see what looks like Anthropic doing work for you, but really it hasn’t been able to bring 100% of its ability to bear. And the gap is pretty big.

Difference Two — No Memory

[04:07] Then the second one is about caching — prompt caching.

[04:09] So what’s prompt caching? Part of why Claude Code is so good is its context — how much of your conversation it can hold. It’s a one-million (1M) token context.

[04:19] So you hand it a project with hundreds of thousands of lines, and if it keeps working on that project, it holds the project in cache — meaning once it’s read through it, the next thing it does, it doesn’t have to read the project again. It sits in the model’s cache and you can call it directly, and it doesn’t charge you extra for it.

[04:32] That has a proper term: prompt caching. And official caching is free for an hour.

[04:37] So say I have it read a 200,000-word novel — within that hour, if I read that novel over and over, as long as the prompt is the same, it won’t charge me, it’s free, and it can be pulled straight from the API.

[04:49] But once you go through a relay, they now attach a random string — which is effectively invalidating that cache. Which means every single time, it re-reads all 200,000 words of that novel. So on the relay side, they can bill you many times over.

[05:02] It looks like 70% off, but in real usage it can end up being the more expensive option.

Difference Three — The Model Gets Swapped

[05:07] The third difference is that the model gets swapped out.

[05:09] If you’re using a relay with no brand or not much reputation, you think you’re connected to the Opus 4.7 model, right — behind the scenes it might be Haiku, or Sonnet, or even some domestic model standing in for your model.

[05:25] Say you run some simple search task — it might just swap the model out, and on your end you have no idea, right? That’s a business-integrity problem.

Data Security + Poisoning

[05:33] There’s one more problem. Never mind poisoning for a second. There’s also the fact that with a lot of relays, everything you send in your conversation can be seen by the relay, complete, in one shot.

[05:46] They might steal your code, or steal your documents — that’s actually the secondary concern. Well, how do I put it, you can’t really call it secondary — they’ll take all of this information of yours and use it for training, or sell it on. So you have no idea that some great idea of yours, or some confidential file, might get exposed or seen by someone in the process.

[06:08] And the poisoning I mentioned means they can slip fake tools into your instructions — it looks like it’s executing the task you wanted, but it’s actually quietly doing other things.

[06:18] How do I even put this — people back home are just too clever.

[06:21] Of course, it’s not that there are no trustworthy relays; some relays are usable. I won’t recommend any here, because for one thing I don’t use relays much myself — when I do use a relay it’s usually for video models or image models, models where there isn’t much private information getting exposed.

The April 4 Cutoff

[06:37] Then on April 4 this year, Anthropic cut off all third-party client access across the board — including tools like OpenClaw and Hermes, the freebie-farming tools. None of them can call from a subscription anymore. So the whole price level went up too.

The Call — Use the Official Subscription

[06:54] So if you’re using Claude Code now, the best way is still an official subscription account.

[07:01] With the $200 plan basically, unless you’re pulling all-nighters back to back building things, you actually can’t max it out.

[07:07] If you do the math at official API rates — if you maxed out that usage, say over these past few months. It basically — I had this happen once, and some of you may remember, Anthropic gave out some credits — so with the $200 plan, say, it would give me $200 of extra usage as credit to spend.

[07:29] At the time I turned on Opus 4.7, and turned on fast mode — and $200 burned down in a bit over 20 minutes. Twenty-odd minutes to burn $200. And even without fast mode, basically an hour, an hour and a bit, and the whole $200 gets drawn down.

[07:48] So again, what we’re saying is — don’t be put off by the hassle, don’t go looking for those so-called easy routes. Right? Let’s take the real official subscription — which as of now I think is the best tool out there — and actually put it to work.

[08:05] That’s what I’ve got for today. See you next episode, bye-bye.