Side note: I don’t know what Anthropic changed but now Claude Code consumes the quota incredibly fast. I have the Max5 plan, and it just consumed about 10% of the session quota in 10 minutes on a single prompt. For $100/month, I have higher expectations.

Relevant: https://www.reddit.com/r/ClaudeAI/comments/1s7zgj0/investiga...

https://www.reddit.com/r/ClaudeAI/comments/1s7mkn3/psa_claud...

That explains things. Im getting this: API Error: 400 {"error":{"message":"Budget has been exceeded! Current cost: 271.29866200000015, Max budget: 200.0","type":"budget_exceeded","param":null,"code":"400"}}

So I completetly ran out of tokens and haven’t even used it at all for the past couple of days, and last week my usage was very light. Let me scratch that, all my usage has been very light since I got this plan at work. It’s a an enterprise subscription I believe, hard to tell since it doesn’t connect directly to Anthropic, rather it goes through a proxy on Azure.

Im not liking this at all and all, so flaky and opaque. Not possible to get a breakdown on what the usage went on, right? Do we have to contact Anthropic for a refund or will they restore the bogus usage?

This is a serious problem with the fact that it's nearly impossible to understand what a "token" is and how to tame their use in a principled way.

It's like if cars didn't advertise MPG, but instead something that could change randomly.

Like if cars measured fuel efficiency or range using the knobs in the tread on your tire.

Relevant post: https://modal.com/blog/dollars-per-token-considered-harmful

(disclaimer: I work with the author)

I completely agree that requests are what should be charged for. But I think there are two things, given that requests aren't all going to cost the same amount:

1. Estimate free invoicing the requests and letting users figure it out after the fact. 2. Somehow estimating cost and telling users how much a request will cost.

We have 1, we want 2.

Also, certain models are more verbose than the others. We are basically at the mercy of a model who likes to ramble a lot.

im fiarly certain the knob on the machine that controls length of redundant comments and docblocks is cranked to 11. it makes me curious how much of their bottom line is driven by redundant comment output.

Anthropic really needs to opensource claude code.

One of the biggest turnoffs as a claude code user is the CC community cargo culting the subreddit because community outreach is otherwise poor.

Looks like your wish was accidentally granted :)

Hilarious. Makes you hope CC team sticks to it, so much goodwill to be had by doing so.

I noticed 1M context window is default and no way not to use it. If your context is at 500-900k tokens every prompt, you’re gonna hit limits fast.

I had to double check that they'd removed the non-1M option, and... WTF? This is what's in `/config` → `model`

    1. Default (recommended)    Opus 4.6 with 1M context · Most capable for complex work
    2. Sonnet                   Sonnet 4.6 · Best for everyday tasks
    3. Sonnet (1M context)      Sonnet 4.6 with 1M context · Billed as extra usage · $3/$15 per Mtok
    4. Haiku                    Haiku 4.5 · Fastest for quick answers

So there's an option to use non-1M Sonnet, but not non-1M Opus?

Except wait, I guess that actually makes sense, because it says Sonnet 1M is billed as extra usage... but also WTF, why is Sonnet 1M billed as extra usage? So Opus 1M is included in Max, but if you want the worse model with that much context, you have to pay extra? Why the heck would anyone do that?

The screen does also say "For other/previous model names, specify with --model", so I assume you can use that to get 200K Opus, but I'm very confused why Anthropic wouldn't include that in the list of options.

What a strange UX decision. I'm not personally annoyed, I just think it's bizarre.

`/model opus` sets it to the original non-1M Opus... for now.

Thanks. I quickly burned through $100 in credit when I started using Opus 4.6 in OpenCode via OpenRouter. My session stopped and was getting an error not representative of credit availability, so was surprised after a few minutes when I finally realized Opus just destroyed those credits on a bullshit reasoning loop it got stuck in. Anthropic seems to know that the expanded context is better for their bottom line as they've defaulted it now.

And as others have said it's very easy to burn token usage on the $100/month plan. It's getting to the point where it's going to very much make sense to do model routing when using coding tooling.

Not sure why you were downvoted because this is actually correct. Can also use --model opus

export CLAUDE_CODE_DISABLE_1M_CONTEXT=1

Anthropic is not building good will as a consumer brand. They've got the best product right now but there's a spring charging behind me ready to launch me into OpenCode as soon as the time is right.

Would you use Opus if you switched to OpenCode?

I'd like to use Opus with OpenCode right now to combine the best TUI agent app with the best LLM. But my understanding is Anthropic will nuke me from orbit if I try that.

You can use Opus with OpenCode anytime you want, just not with the Claude plan. You can use it via API with any provider, including Anthropic's API. You can use it with Github Copilot's plan. The only thing you can't do without getting banned is use OpenCode with one of Claude's plans.

If you want to use Opus with a different coding harness along with a coding plan, you can use Github CoPilot. It even has built in authentication with OpenCode.

OpenCode with a Copilot Business sub and Opus 4.6 as the model works well

I'm looking at their plans (https://github.com/features/copilot/plans) it seems like the limits might be pretty low, even with the Pro+ plan which is 2x the cost of Claude Pro. It seems like Claude Pro might be 10-20x the Opus tokens for only twice the price.

Copilot has a totally different billing model. It's request based rather than token based. Counter-intuitively, in our case at least, it is way cheaper than token based pricing. One request can sometimes consume 2-4 million tokens but is billed as a single request (or it's multiplier if using a premium model like opus).

do you pay for the full context every prompt? what happened with the idea of caching the context server side?

You don't. Most of the time (after the first prompt following a compaction or context clear) the context prefix is cached, and you pay something like 10% of the cost for cached tokens. But your total cost is still roughly the area under a line with positive slope. So increases quadratically with context length.

It helps a ton but it doesn't last forever and you still have to pay to write to the cache

I've been jumping from Claude -> Gemini -> GPT Codex. Both Claude and Gemini really reduced quotas and so I cancelled. Only subbed GPT for the special 2x quota in March and now my allocation is done as well.

I decided to give opencode go a try today. It's $5 for the first month. Didn't get much success with Kimi K2, overly chatty, built too complex solutions - burned 40% of my allocation and nothing worked. ¯\_(ツ)_/¯.

But Minimax m2.7. Wow, it feels just like Claude Opus 4.6. Really has serious chops in Rust.

Tomorrow/Wednesday will try a month of their $40 plan and see how it goes.

Minimax 2.7 is great. Not close to Claude but good enough for a lot of coding tasks.

GLM-5 (and 5.1) is surprisingly impressive too I’m finding.

They need to get to profitability because that sweet sweet Saudi subsidy cash is gone gone.

They wont be profitable at this point...they just dont realise they are eating their own tail.

I've heard this a few times lately, but this past weekend I built a website for a friend's birthday, and it took me several hours and many queries to get through my regular paid plan. I just use default settings (Sonnet 4.6, medium effort, thinking on).

I'm guessing Opus eats up usage much, much faster. I don't know what's going on, since a lot of people are hitting limits and I don't seem to be.

Update: Maybe the difference is that I think I was just using the vscode extension at the time: https://news.ycombinator.com/item?id=47586176

I go back and forth between vscode and claude in the terminal, but that day I think I did vscode.

what they changed was peak vs off-peak usage metering.

using it on the weekend gets you more use than during weekdays 9-5 in US eastern time.

I waited until off peak hours to use Opus 4.6 to do some research. One prompt consumed 100% of my 5h limit and 15% of my weekly usage. Even off peak it's still insane. Opus didn't even manage to finish what it was doing.

I'm surprised it's during east coast working hours and not west coast.

the speculation i read was that it's trading hours, and they're getting a lot of load from the finance industry

Technically, this was Friday morning, so I think I was still in peak hours.

Even with Opus I don’t usually hit limits on the standard plan. But I am not doing professional work at the moment and I actually alternate between using the LLM and reading/writing code the old fashioned way. I can see how you’d blow through the quota quickly if you try to use LLMs as universal problem solvers.

Have had similar issues with costs sometimes being all over the map. I suspect that the major providers will figure this out as it’s an important consideration in the enterprise setting

This is a very normal thing to be the top comment on an article on how to use Claude Code.

I'm very surprised to see enshittification starting so early. I was expecting at last 3-4 years of VC subsidized gravy train.

This has been 6 months of constant decline so at this point I am wondering when they cliff it like wework

Looks like they are falling victim to their own slop. This smells a lot like the Amazon outages caused by mandated clanker usage.

Reminds me of when I would mess with my friends on "pay per text" plans by sending them 10 text messages instead of just 1. I should start paying attention to unattended laptops and blow up some token usage in the same manner.

It's almost like an evolution of bobby tables.

Since it’s an 1.44M image I assume they use 3.5” diskettes. The terms floppy and diskette are used as synonyms today, but the different names make sense since floppies are flexible and “floppy”. Diskettinux?

Docker won’t save you from prompt injektions that attack your network.

No kidding? https://taoofmac.com/space/blog/2026/01/12/1830

Still, I don’t think bubblewrap is either a simple or safe enough solution.

It won’t save you from prompt injektions that attack your network.

Shameless plug, in case you're interested: https://github.com/EstebanForge/construct-cli

Let me know if you give it a go ;)

Interesting, any plans to add LiteLLM (https://github.com/BerriAI/litellm) and Kilocode (https://github.com/Kilo-Org/kilocode)?

Added https://github.com/EstebanForge/construct-cli/releases

Will check those out :)

In theory the docker container should only have the projects directory mounted, open access to the internet, and thats it. No access to anything else on the host or the local network.

Internet to connect with the provider, install packages, and search.

It's not perfect but it's a start.

Docker containers run in their separate isolated network

of course, I'm not pretending this is a universal remedy solving all the problems. But I will add a note in the readme to make it clear, thanks for the feedback!

Let’s not forget Amiga 3000UX - a workstation released with Amiga Unix, a full port of AT&T Unix System V Release 4 (SVR4). Notable users include Free Software Foundation staff programmers who used it at MIT to help further some early development of the GNU operating system.

https://en.m.wikipedia.org/wiki/Amiga_3000UX

The 3000UX was dead on arrival

It cost as much as an early 1990s UNIX workstation but it featured the technology of the 1980s, so it was extremely slow by the standards of the day.

For the price of a 3000UX, you could buy an SGI with 10x the CPU power, or a Sun with 10x as many pixels on the display. It was a really, really bad deal. As per usual for Commodore, too little, too late.

This is very annoying - I get this too now, and there is no way to make it go away (except for enabling watch history I assume). Very bad!

> - What is a belief you had as a child that you no longer have? ...

Hm, the type of some of these questions resemble the type of the personal questions used for password recovery by some companies. As a paranoid person, I am reluctant to disclose this information to unknown people.

But you should never give the correct answer for password recovery.

What is a belief you had as a child that you no longer have?

Purple Ocean.

And that can be your answer for all the questions: first pet name, elementary school, and so on.

> But you should never give the correct answer for password recovery.

Exactly. You're supposed to use your password manager to generate a second password and use that as your answer. I know this sounds stupid but it is the only way to stay safe.

The only problem with this is if you have to read it to someone live. Otherwise, yes!

i've heard stories of call centers accepting "oh, i don't remember, i just typed a bunch of random letters and numbers" as confirmation over the phone.

Very good idea!

Good idea!

You already started the discussion :)

Ah, for me it was cosmic-deflect-attitude. I really struggled with it in my teenage years.

How could that possibly be a password recovery question? A password recovery question should only have one correct answer.

Yep. Not to mention sharing your true opinions on deep/intimate topics to strangers is generally a terrible idea especially on camera.

Also, these types of posts make me wonder what the goals of the project are. Is the intent here to gather data and sell it? Gather users and sell the company? I don't believe in saints. I don't believe in the "don't be evil" mantra.

For all the talk about privacy and anonymity, seems many here want to give away their privacy and anonymity.

Have you tried applying for a position in a public sector IT-department? The public sector should not age-discriminate, at least in theory.

I think he was great in The Emigrants and The New Land.

Sounds more like an index of beginner languages.

"The PYPL PopularitY of Programming Language Index is created by analyzing how often language tutorials are searched on Google."