Have been using this for awhile, I'm on the $100/mo Max plan and have been running $600-800/mo in terms of usage, and I'm hardly pushing it to the limits (missing lots of billing windows).
It makes me wonder what Anthropic's true margins are. I could believe they are overcharging via the API, Sonnet is $3/$15/Mtok and Opus at an ABSURD $15/$75/Mtok. But to break even for me, that would mean that they're overcharging by a factor 5x-10x, which doesn't seem possible. Is the music going to stop for Claude Code the same way it did for Cursor? I have to imagine every incentive in the world is pushing them to lower inference cost rather than introduce stricter limits, and unlike Cursor they can actually can reach into their stack and do this. But I'm not sure they're capable of miracles.
Regardless, I'm bullish Anthropic. Sonnet and Opus don't benchmark as well as O3/Grok4 at pure coding, and aren't as cheap as Kimi K2 for theoretically similar perf, but as any user knows they are top tier at instruction following, highly reliable and predictable, and have a certain intangible theory of mind that is unique to Anthropic.
> but as any user knows they are top tier at instruction following, highly reliable and predictable
This is spot on. Reliability is really the #1 priority for me when it comes to coding agents, and Sonnet, and especially Opus, really deliver on it. It makes such a huge difference when it comes to agents. Anthropic really nailed it on this.
My process has become: get Opus to generate a plan, use o3 to help me review the plan, and then get Opus to implement the plan. This works extremely well for me, and is the first time where I've felt AI being actually useful for coding anything more than small prototypes.
> but as any user knows they are top tier at instruction following, highly reliable and predictable, and have a certain intangible theory of mind that is unique to Anthropic.
I’m not such an “any user” as in my experience they are exceptionally terrible at following my instructions and cutting edges whenever they like, I have to iteratively get more and more specific at what they should not do, all the time trying to predict their misbehavior. To then see a certain point or a threshold at which they throw my instructions out the window to match and reproduce the very things or patterns I wanted to avoid and likewise with the things and patterns I explicitly instructed them to use or apply.
Very unreliable.
> have a certain intangible theory of mind that is unique to Anthropic.
What do you mean by music stopping for Cursor? Almost every single developer I run into is transitioning/transitioned to it today. It's stinks like the new VS Code.
There's a predictable journey: people start with Cursor when they are new to AI, and quickly move on to something more powerful once they realise that the IDE [1] is holding em back and that forking VSCode is [2][3] tech-debt.
> Sonnet and Opus don't benchmark as well as O3/Grok4 at pure coding
Do any of the others have a "claude code" local agent? Seems like a big gap IMO. Though, it should be pretty easy for them to close that gap.
I don't usually take too many moral stances but I feel like I can't use Grok. It's bad enough Musk did his Nazi salute but his AI product itself is a Nazi too? It might be good at coding but I really can't stomach using it.
FWIW, people report that Grok 4 is not very good at coding, and xAI admit this themselves when they said they will be releasing a separate coding model in "the next few weeks".
Also, Google does have Gemini CLI, OpenAI does have Codex CLI, and then there is Aider which can support any model. I think the big difference is that Anthropic's models are the best for this use-case right now, and Anthropic has the Max plan which makes a massive difference to the cost of using Claude Code compared to competitors (although the Gemini CLI has insane free tiers).
I'm not sure how this will play out in the future, because it seems to me that Claude Code does not have much of a moat beyond Anthropic having the best coding models right now, and them offering model usage at heavily discounted prices.
I used it really heavily for a few days, but now I just don’t have the time to send much instructions. Still used on the order of $1000 in the first month, but imagine it will go down as time goes on. CC is so convenient I think I’d have a hard time living without it though
I doubt that there are many inactive subscribers, as Claude Code / Max Plan is relatively new. They might be hoping that way in a couple of months, though.
Exactly, swyx. Any flat rate pricing plan is effectively a bet against the future. It's a grab for engineers that's subsidised. Now, the problem is that GPUs are expensive; they are a costly resource to use. Inferencing is expensive.
So what happens is inevitable:
- Wild promises of unlimited usage and consumers feeling tricked when the impossible is impossible to deliver (Cursor pricing changes).
- Quasi-unlimited usage with rate-caps, but the models get quantised to all hell? [search Twitter for folks reporting Claude feels dumber around/near outages].
- Engineers sharing tools and techniques on how to squeeze pounds out of a flat-rate plan (original post), which results in more power users doing that, which puts more pressure on margins.
I started on the Pro plan a week ago and was already contemplating jumping to Max. When I hit a limit yesterday I upgraded to Max and hit a limit again before seeing the news of the changed usage limit.
For what it's worth, everything seems fixed today.
People have been saying this since it first came out. I don’t doubt there are occasional bugs/service disruptions but personally I really doubt Anthropic is silently decreasing the limits.
I really like how easy it is to run using bunx, pnpx, npx, etc.
But does anyone have thoughts on the security aspect. Getting people used to just running code like this that has full access to the system is slightly concerning.
On the other hand it’s no different than installing npm packages
Yes, you need to run these agents in a sandboxed environment when running full AFK [1] yolo. That could be a Docker container or it could be remote developer environment.
Maybe this kind of thing would be better written in Deno?
Deno has mechanisms for allow-listing the exact files the process can access - in this case you would want to give it read-only access to the log files in the ~/.claude directory and nothing else.
Have been using this for awhile, I'm on the $100/mo Max plan and have been running $600-800/mo in terms of usage, and I'm hardly pushing it to the limits (missing lots of billing windows).
It makes me wonder what Anthropic's true margins are. I could believe they are overcharging via the API, Sonnet is $3/$15/Mtok and Opus at an ABSURD $15/$75/Mtok. But to break even for me, that would mean that they're overcharging by a factor 5x-10x, which doesn't seem possible. Is the music going to stop for Claude Code the same way it did for Cursor? I have to imagine every incentive in the world is pushing them to lower inference cost rather than introduce stricter limits, and unlike Cursor they can actually can reach into their stack and do this. But I'm not sure they're capable of miracles.
Regardless, I'm bullish Anthropic. Sonnet and Opus don't benchmark as well as O3/Grok4 at pure coding, and aren't as cheap as Kimi K2 for theoretically similar perf, but as any user knows they are top tier at instruction following, highly reliable and predictable, and have a certain intangible theory of mind that is unique to Anthropic.
> but as any user knows they are top tier at instruction following, highly reliable and predictable
This is spot on. Reliability is really the #1 priority for me when it comes to coding agents, and Sonnet, and especially Opus, really deliver on it. It makes such a huge difference when it comes to agents. Anthropic really nailed it on this.
My process has become: get Opus to generate a plan, use o3 to help me review the plan, and then get Opus to implement the plan. This works extremely well for me, and is the first time where I've felt AI being actually useful for coding anything more than small prototypes.
How do you switch to o3 to review the plan?
I have a workflow that tells Claude Code to generate a planning markdown document: https://gist.github.com/Sothatsit/c9fcbcb50445ebb6f367b0a6ca...
Personally, I use Repoprompt for this + deeper context integration.
I use zenMCP to collaborate with opus
Have CC output a plan.md file.
> but as any user knows they are top tier at instruction following, highly reliable and predictable, and have a certain intangible theory of mind that is unique to Anthropic.
I’m not such an “any user” as in my experience they are exceptionally terrible at following my instructions and cutting edges whenever they like, I have to iteratively get more and more specific at what they should not do, all the time trying to predict their misbehavior. To then see a certain point or a threshold at which they throw my instructions out the window to match and reproduce the very things or patterns I wanted to avoid and likewise with the things and patterns I explicitly instructed them to use or apply.
Very unreliable.
> have a certain intangible theory of mind that is unique to Anthropic.
That’s quite poetic.
What do you mean by music stopping for Cursor? Almost every single developer I run into is transitioning/transitioned to it today. It's stinks like the new VS Code.
Their pricing change recently has people reaching for alternatives.
There's a predictable journey: people start with Cursor when they are new to AI, and quickly move on to something more powerful once they realise that the IDE [1] is holding em back and that forking VSCode is [2][3] tech-debt.
[1] https://ghuntley.com/overton
[2] https://ghuntley.com/fracture
[3] https://ghuntley.com/amazon-kiro-source-code/
> Sonnet and Opus don't benchmark as well as O3/Grok4 at pure coding
Do any of the others have a "claude code" local agent? Seems like a big gap IMO. Though, it should be pretty easy for them to close that gap.
I don't usually take too many moral stances but I feel like I can't use Grok. It's bad enough Musk did his Nazi salute but his AI product itself is a Nazi too? It might be good at coding but I really can't stomach using it.
FWIW, people report that Grok 4 is not very good at coding, and xAI admit this themselves when they said they will be releasing a separate coding model in "the next few weeks".
Also, Google does have Gemini CLI, OpenAI does have Codex CLI, and then there is Aider which can support any model. I think the big difference is that Anthropic's models are the best for this use-case right now, and Anthropic has the Max plan which makes a massive difference to the cost of using Claude Code compared to competitors (although the Gemini CLI has insane free tiers).
I'm not sure how this will play out in the future, because it seems to me that Claude Code does not have much of a moat beyond Anthropic having the best coding models right now, and them offering model usage at heavily discounted prices.
> people report that Grok 4 is not very good at coding
There are agentic models and oracle models. It can be modelled on a four-way quadrant of agent vs oracle and high safety vs low safety.
https://ghuntley.com/cars
Grok is oracle and low safety.
Grok4 is pretty decent at planning and figuring out libraries and APIs.
For code it falls down past simple scripts and utilities.
perhaps there are enough inactive subscribers to compensate for the heavy users.
I used it really heavily for a few days, but now I just don’t have the time to send much instructions. Still used on the order of $1000 in the first month, but imagine it will go down as time goes on. CC is so convenient I think I’d have a hard time living without it though
I doubt that there are many inactive subscribers, as Claude Code / Max Plan is relatively new. They might be hoping that way in a couple of months, though.
any dropoff in usage limits for you recently? https://techcrunch.com/2025/07/17/anthropic-tightens-usage-l...
cant help but think that you guys yelling about it so loudly from the rooftops is really really not helping your case lol
Exactly, swyx. Any flat rate pricing plan is effectively a bet against the future. It's a grab for engineers that's subsidised. Now, the problem is that GPUs are expensive; they are a costly resource to use. Inferencing is expensive.
So what happens is inevitable:
- Wild promises of unlimited usage and consumers feeling tricked when the impossible is impossible to deliver (Cursor pricing changes).
- Quasi-unlimited usage with rate-caps, but the models get quantised to all hell? [search Twitter for folks reporting Claude feels dumber around/near outages].
- Engineers sharing tools and techniques on how to squeeze pounds out of a flat-rate plan (original post), which results in more power users doing that, which puts more pressure on margins.
In goose meme format, "What are the margins?"
https://x.com/GeoffreyHuntley/status/1945636266009399414
> Any flat rate pricing plan is effectively a bet against the future
How quickly we forget Moore's law, or at least what has replaced Moore's law.
I started on the Pro plan a week ago and was already contemplating jumping to Max. When I hit a limit yesterday I upgraded to Max and hit a limit again before seeing the news of the changed usage limit.
For what it's worth, everything seems fixed today.
People have been saying this since it first came out. I don’t doubt there are occasional bugs/service disruptions but personally I really doubt Anthropic is silently decreasing the limits.
I'm the author of ccusage. Thank you for sharing! Also please refer our website. https://ccusage.com
I really like how easy it is to run using bunx, pnpx, npx, etc.
But does anyone have thoughts on the security aspect. Getting people used to just running code like this that has full access to the system is slightly concerning.
On the other hand it’s no different than installing npm packages
> anyone have thoughts on the security aspect
Yes, you need to run these agents in a sandboxed environment when running full AFK [1] yolo. That could be a Docker container or it could be remote developer environment.
[1] https://ghuntley.com/ralph
They are talking about the fact that you can run this npm tool without installing it - not running code agents
Ah, that's the least of their issues!
Maybe this kind of thing would be better written in Deno?
Deno has mechanisms for allow-listing the exact files the process can access - in this case you would want to give it read-only access to the log files in the ~/.claude directory and nothing else.
[flagged]
You’ve pasted this exact comment 13+ times. Could you please stop spamming?