Comment by dfabulich

Well OpenCode already exists and you can connect it to multiple providers, so you could just say that the agentic CLI harness business model as a service/billable feature is no more. In hindsight I would say it never made sense in the first place.

Nah, I think Opus is fantastic but not Claude Code. Their models are way better.

Claude code is nothing more than a loop to Opus.

> the reality is that Claude Code is what makes Anthropic so relevant in the programming more, much more than the Claude models themselves

but Claude Code cannot run without Claude models? What do you mean?

Yeah, I've heard of people swapping out the model that Claude Code calls and apparently its not THAT much of a difference. What I'd love to see from Anthropic instead is, give me smaller LLM models, I don't even care if they're "open source" or not, just pull down a model that takes maybe 4 or 6 GB of VRAM into my local box, and use those for your coding agents, you can direct it and guide it with Opus anyway, so why not cut down on costs for everyone (consumer and Anthropic themselves!) by just letting users who can run some of the compute locally. I've got about 16GB of VRAM I can juice out of my Macbook Pro, I'm okay running a few smaller models locally with the guiding hand of Opus or Sonnet for less compute on the API front.

It's probably not a "loss-leader" so much as "somewhat lower margin". Their bizdev guys are doubtless happy to make a switch between lower-margin, higher-multiple recurring revenue versus higher-margin, lower-multiple pay-as-you-go API billing. Corporate customers with contracts doubtless aren't paying like that for the API either. This is not uncommon.

When you have a "loss leader" whose sole purpose is to build up market share (e.g. put competitors out of business) that's called predatory pricing.

My guess is that ultimately the use of Claude code will provide the training data to make most of what you do now in Claude code irrelevant.

I keep hearing those claims that's they lose money on it, but I have more and more doubts about this being true.

GPU compute cost has falled down in the last two years a lot.

Do you think Anthropic followed all the ToS of every website on the internet when scraping them for training data?

Poor behavior is still poor behavior even if the relevant ToS aligns with it.

I can easily get Claude Code to run for 8-10 hours unsupervised without stopping with sub-agents entirely within Claude Code.

I think it is more likely that if you stick with Claude Code, then you are more likely to stick with Opus/Sonnet, whereas if you use a third party CLI you might be more likely to mix and match or switch away entirely. It's in their interest to get you invested in their tooling.

> I guess one issue is that you pay $200/month whether you use it or not.

I can easily churn through $100 in an 8 hour work day with API billing. $200/month seems like an incredibly good deal, even if they apply some throttling.

Why is supervising one's agents to be as efficient as possible a problem for Anthropic?

What kind of ways to supervise?

Another analogy is that it’s a takeout but anthropic is insisting you only eat at home with the plastic utensils they’ve provided rather than the nice metal utensils you have at home.

Another analogy is that it’s a restaurant that offers delivery and they’re insisting you use their own in house delivery service instead of placing a pickup order and asking your friendly neighbor to pick it up for you on their way back from the office.

It's not really a fair analogy. Restaurants don't want you taking food away because they want to limit the amount you eat to a single meal, knowing that you'll stop when you get full. If you take food out you can eat more by waiting until the next meal when you're hungry again.

You don't "get full" and "get hungry again" by switching UIs. You can consume the same amount whether you switch or you don't switch.

Not really. At a buffet restaurant, if you could take the food out with you, you'd takeaway more food than you can eat at one sitting. OpenCode users and Claud Code™ CLI users use tokens at approximately the same rate.

This is more like an all-you-can-eat restaurant requiring you to eat with their flimsy plastic forks, forbidding you to bring your own utensils.

...no, that's more like "but you can't bring your own fork"

anthropic should not be criticizing the gluttony of others whilst licking its fingers surrounded by buckets full of fried chicken

I’d bet almost everyone who opts to buy the $200 plan is happy with the deal they’re getting relative to API pricing.

I think some people get triggered by the inconsistency in pricing or the idea of having a fixed cost for somewhat vague usage limits.

In practice it’s a great deal for anyone using the model at that level.

Even at the Max plan it's not unlimited. IIRC they say the limit is something like 20x the $20 plan.

With "normal use" you're unlikely to hit the limit though, unless you only use the Opus model and do things in parallell.

It is not unlimited, being not careful with your context management, you hit the limits quickly.

As far as I know, OpenCode sends (has to send) the same data to Anthropic as Claude Code™ CLI (especially if they're going to successfully imitate CC™ in order to access cheap subscription pricing).

Claude Code only trains on data if you opt in

That is not true, though. You have to opt in for them to train on your data

Claude code cli and api vs subscription are tangential. You can use Claude code with an api token.

In some sense that's what OpenCode is, and Anthropic's not having that.

I suspect that they lose control over the cheaper models CC can choose for you for eg. file summaries or web fetch. Indeed, they lose web fetch and whatever telemetry it gives them completely.

It's not unreasonable to assume that without the ability to push haiku use aggresively for summarization, the average user in OC vs CC costs more.

Not that hard to understand, they want to control how their users use their product. A CLI they built, even acquiring the framework it was built in, is a way to achieve that.

It's because the model companies believe there's no way to survive just selling a model via an API. That is becoming a low margin, undifferentiated commodity business that can't yield the large revenue streams they need to justify the investments. The differences between models just aren't large enough and the practice of giving model weights away for free (dumping) is killing that business.

So they all want to be product companies. OpenAI is able to keep raising crazy amounts of capital because they're a product company and the API is a sideshow. Anthropic got squeezed because Altman launched ChatGPT first for free and immediately claimed the entire market, meaning Anthropic became an undifferentiated Bing-like also-ran until the moment they launched Claude Code and had something unique. For consumer use Claude still languishes but when it comes to coding and the enormous consumption programmers rack up, OpenAI is the one cloning Claude Code rather than the other way around.

For Claude Code to be worth anything to Anthropic's investors it must be a product and not just an API pricing tier. If it's a product they have so many more options. They can e.g. include ads, charge for corporate SSO integrations, charge extra for more features, add social features... I'm sure they have a thousand ideas, all of which require controlling the user interface and product surface.

That's the entire reason they're willing to engage in their own market dumping by underpricing tokens when consumed via their CLI/web tooling: build up product loyalty that can then be leveraged into further revenue streams beyond paying for tokens. That strategy doesn't work if anyone can just emulate the Claude Code CLI at the wire level. It'd mean Anthropic buys market share for their own competitors.

N.B. this kind of situation is super common in the tech industry. If you've ever looked at Google's properties you'll discover they're all locked behind Javascript challenges that verify you're using a real web browser. The features and pricing of the APIs is usually very different to what consumers can access via their web browser and technical tricks are used to segment that market. That's why SERP scraping is a SaaS (it's far too hard to do directly yourself at scale, has to be outsourced now), and why Google is suing them for bypassing "SearchGuard", which appears to just be BotGuard rebranded. I designed the first version of BotGuard and the reason they use it on every surface now, and not just for antispam, is because businesses require the ability to segment API traffic that might be generated by competitors from end user/human traffic generated by their own products.

If Anthropic want to continue with this strategy they'll need to do the same thing. They'll need to build what is effectively an anti-abuse team similar to the BotGuard team at Google or the VAC team at Valve, people specialized in client integrity techniques and who have experience in detecting emulators over the network.

DAU/MAU for IPO.

I see zero reason to believe the $200 subscription is losing money. Anthropic makes subscriptions cheaper because 1. Most users dont use all their allocated tokens 2. subscriptions create a lock in effect, even if its a weak one 3. Easier to raise money when you can point to your ARR from subscriptions 4. Lowering revenue variance month to month is very valuable for businesses

If true I wonder what kind of feedback loop is happening by training on human behavior that's directly influenced by the output of the same model

They don't exactly put it front and center. Click "usage limits" (at the bottom; https://support.claude.com/en/articles/9797557-usage-limit-b...) then "Max plan" in the first list (https://support.claude.com/en/articles/11014257-about-claude...). There is a $200/mo price which people are likely referring to with "20x more usage per session" (which kinda bothers me because I'd bet my bottom dollar it's 20x "as much" but that's a lost cause).

I got a pro subscription yesterday. With it you get a certain amount of tokens and you have a certain limit every 5 hours and every week.

Once the limit is reached, you can choose to pay-per-token, upgrade your plan, or just wait until it refreshes. The more expensive subscription variants just contain more tokens, that’s all.

Keep scrolling down, there is a Max option

Not the OP, but it seems pretty clear to me - they're suggesting that fixed per-month pricing with unlimited usage shouldn't exist at all, as it doesn't really make sense for a product that has per token costs.

Instead, they're saying that a 200$/month subscription should pay for something like $250 worth of pay-per-token API tokens, and additionally give preferential pricing for using up more tokens than that.

So, if the normal API pricing were 10$ per million tokens, a 200$ per month subscription should offer 25M tokens for free, and allow you to use more tokens at a 9$/1M token rate. This way, if you used 50M tokens in a month, you'd pay 445$ with a subscription, versus paying 500$ with pay-as-you-go. This is still a good discount, but doesn't create perverse incentives, as the current model does.

I think they are.

The key question is about why they want to you to use the CLI. If you're not the customer, you're the product.

There's also a monopolistic aspect to this. Having the best model isn't something over can legally exploit to gain advantage in adjacent markets.

It reeks of "Windows isn't done until Lotus won't run," Windows showing spurious error messages for DR-DOS, and Borland C++ losing to the then-inferior Visual C++ due to late support of new Windows features. And Internet Explorer bundling versus Netscape.

Yes, Microsoft badly wanted you to use Office, Visual C++, MS-DOS, and IE, but using Windows to get that was illegal.

Microsoft lost in court, paid a nominal fine, and executives were crying all the way to the bank.

Well they are doing the same to website owners who rely on human visitors for their revenue streams.

Both scraping and on-demand agent-driven interactions erode that. So you could look at people doing the same to them as a sort of poetic justice, from a purely moral standpoint at least.

Assuming the actual price for many user is closer to 1k USD/mth than to 200 USD/mth, and the actual price is closer to their target margin to be a viable business, they're practically subsidising usage after 200 USD/mth. Together with other AI-TECH doing the same, they fabricate a false sense of "AI is capable AND affordable", which imo is evil.

it doesn't need sudo privileges...

The price is the same whether you run it on Bedrock or Anthropic....

Unless I'm severely mistaken that's not the source code for claude-code. It's a few official plugins and some helper scripts

Look a bit closer to the contents of this repo. There is basically no code.

Right? Lots of "Anthropic should do x because that's what makes sense to me".

For the bizarre amount of work that gets done for that 180 euro, it is really cheap. We are just getting used to it and sinking prices everywhere, it is just that CC is the best (might be taste or bias, I at least think so), so we are staying with it for now. If it gets more expensive, we will go and try others for production instead of just trying them to get a feel for the competition as we do now.

This take is ridiculous. Nearly everyone who uses Max agrees that what they get for the money paid is an amazing deal. If you don't use or understand how LLMs fit in your workflows, you are not the target customer. But for people who use it daily, it is a relatively small investment compared to the time saved.

That entirely depends on your business case. If that call costing 50 Cent has done something for me which would have taken me more than 1 minute of paid working time to do it's sustainable.

It pays for itself in a day for some folks. It is a lot but it’s still cheap.

If you're going to link a repository, you should read it first. That repository is just a couple plugins and community links. Claude Code is, and always has been, completely closed source.

That repo does not contain the source code for Claude Code.

You can't use it as an SDK though, unlike codex