Hacker News Reader

Comment by abtinf | Hacker News Reader

That tweet only makes things worse. On top of all their other nonsense recently, it actually convinced me to cancel my subscription.

I can't trust Anthropic to manage their products in a way that supports my workflow.

View on HN

pretty much none of these big providers are offering the guarantees needed to be taken seriously in workplaces right now. the technology itself isn't offering the deterministic guarantees that should warrant it in the workplace right now. problem is everyone's foot is just on the gas. even if your workplace isnt paying for it, people are just straight up rolling their own personal claude accounts to do work at orgs.

ive been trying to make the case all year that if we're going to let employees do shit with ai, lets try claude. in the past like.. 2-3 weeks all that goodwill has basically evaporated.

local inference needs to take off asap because all of these entities actually suck and i wouldn't trust a single sla with anthropic. they are not acting like a serious company right now, this is a joke.

Anthropic is absolutely taken seriously in workplaces, what are you even talking about?

No serious business uses Pro or Max, they are all on Anthropic API billing.

In fact with this move it is plainly obvious that Anthropic is moving compute from prosumers towards enterprise.

But the API is incredibly expensive. I calculated that I would have spent 3000 EUR the last month, a lot more than the 100 I pay now.

Nothing for large companies though.

I am pretty sure that a hole in the pocket in the order of 50 000 000 USD/month (assuming around 20 000 people using AI in not the smartest or most optimized way possible, therefore burning A LOT of tokens) will be noticeable by even the largest companies.

It is noticeable and even promoted, large companies do pay such sums for the API, like $5k+ per person per month. Not every eng is using AI that much already, but companies are clearly willing to pay those sums.

Per employee?

This not nothing.

I know of a very serious business that deployed Max to all of their developers. API pricing, from what I see, can become more expensive than just hiring another dev.

Well yes it is expensive, but companies are paying for that. It is far more expensive than the Max and it does go up to or more in some cases compared to the employee salary.

Larger companies are using Claude through AWS Bedrock and are willing to easily pay $5k+ per engineer per month for it.

The thinking appears to be that a model that can do the work of a developer must be worth a significant share of a developer salary. I think this idea is flawed.

Developer salaries are driven up by scarcity - scarcity of developer skills overall and scarcity of developer skills in specific places like California. If AI models destroy the scarcity then the price worth paying for a coding agent will drop dramatically.

Maybe Anthropic can get away with it for a couple of months. But this will not last.

But if e.g. a developer can do 50% more, shouldn't it be worth it to pay up to 50% of developer salary for the product?

So the % is debatable of course. There's cases where an AI agent can save weeks worth of investigation, there's cases where you are mainly blocked due to processes, and many different circumstances. It's up to every company on their own to decide it. But if they decide it's 50%, why shouldn't they spend 50% of salary on it?

Like imagine a large company with thousands of microservices. You need to build a feature, before you had to setup cross timezone team meetings to figure out who owns what, what is happening in each microservice, how it all connects together. But now you can essentially send an AI Agent to scour and prepare all this material for you, which theoretically in this planning could save hours of back and forth meetings.

If 1 hour / 1 eng costs $200, then a 10 people 1h meeting avoided would save $200 x 10 = $2000 alone.

I don't see it as a replacement for dev, it's more of a multiplier.

I believe what GP is saying is that there is a price calculation today, but then if enough devs become unemployed, their salary will go down, making them more competitive by finops calculations, at which point the Ai prices will have to come down as well. Where equilibrium is, no one knows

I think it's an interest hypothesis but I don't think it works out like that. AI prices aren't priced in relation to the work they do, they're priced in relation to tokens (input/output). As long as it's cheaper to use those tokens than it is to pay a dev, then dev salaries will likely fall. Whenever it becomes cheaper to hire a dev than to use AI, a company will likely just hire a dev. But AI prices won't fall just because dev salaries have.

Yeah, I mean I think there's just too much work and I think devs who are effective with AI won't become unemployed, but their productivity will be multiplied. More will be expected of companies in terms of output, so it will be just more output.

>But if e.g. a developer can do 50% more, shouldn't it be worth it to pay up to 50% of developer salary for the product?

That's the upper bound but it's not the market price.

Accounting software (+ hardware) doesn't cost nearly as much as the accountant hours it saves. Accountant salaries are simply not a relevant yardstick for the price that software vendors can charge for accounting software.

Equally, the market price for code generators will not stay anywhere near the price of developer hours it saves. It will be determined by competition.

Because accounting software is cheaper due to competition. In software eng Claude is currently strongest and there's higher costs involved than normal SaaS. There are many fields in which the tools/machinery cost more than the salaries of people.

>Because accounting software is cheaper due to competition. In software eng Claude is currently strongest and there's higher costs involved than normal SaaS.

Yes, competition not salaries determines the margins that software vendors can charge. That's exactly what I'm saying.

My expectation is that competition between coding agents will stay strong and costs for the current level of software engineering performance will fall.

>There are many fields in which the tools/machinery cost more than the salaries of people.

For example?

Agriculture, oil drilling, trucking, etc...

The machines in those sectors do not cost nearly as much as doing the same work manually. Not even close.

Of course not, but they cost more than the person using them, it's multiplying the productivity of that person, so if AI multiplied enough as well it would make sense.

That's beside the point.

The question was whether the salaries that would have been paid for the working hours replaced by the machine are a realistic yardstick for the market price of the machine.

That's clearly not the case in any industry.

In my calculation a good opus developer can do 10x more, not just 50%.

Got all my tickets from the last two years fixed on a few days. And implemented all the ideas which came to my head.

We're also not seeing much difference in real throughput at an agency. Everyone is getting decent results, output wise but it just doesn't seem to change the outcomes that much. There is also a mixed incentive at an agency, because a reduction in hours spent is a reduction in revenue.

It will be interesting to see how it all plays out, but I suspect if cost continues to increase and output only improves incrementally from here, that the cost will be the final decider rather than the competence.

I could see it being a thing we use only sometimes, for some things, but ultimately remain reliant on developers to get the work through the pipeline.

API usage is on-demand, employees are a constant cost, guess what management loves most.

That's true but employees offer more than code output, and you still need people operating the "machine" at this stage.

I am interested in how corporate politics evolves in this new environment. Usually all the way up the chain, managers and directors use head count as a measure of power and influence (and compensation). Who's going to pay a director top level pay when all they're doing is funneling requirements to various agents? That seems like a technical role that isn't particularly aligned with the soft skills management excel with either.

Management seems disconnected from reality. Real employees accumulate tribal knowledge, have an almost infinite context, and don’t keep disabling unit tests because they don’t pass. They don’t really cost money if it’s information workers that build almost all of the modern service industry. It’s management that we should see as a cost center.

I work for a "very serious" company with many billions of dollars of revenue. All our SWEs have max subscriptions.

My startup just rolled out MAX plans for our engineers. we literally jumped into AI last week. Been in operation for 6 years and are profitable.

What are you guys subscribed to if not Claude? Copilot? Or is everyone legit bringing their own license?

I have the basic subscriptions for Copilot, Claude and Codex. About €50 per month.

I enjoy Codex the most

But like Claude I’m not loyal to any of them.

Copilot's per-prompt pricing model is overwhelmingly the best value for money right now, although they more than doubled the price of Opus 4.7 compared to 4.6 and completely removed 4.5 and 4.6, which erodes their lead somewhat. Copilot restricts context much more than CC, but I still find it to be plenty capable. I've occasionally managed to give Copilot/Opus 4.6 a prompt that kept it productive for a full work day for a cost of just ~$0.10.

I got a claude api key from work that I use with zed. It works really well and usually ends up costing $25/day if I use it all day.

I did recently experiment with copilot for personal stuff, but will cancel it now that opus is no longer available on the $10/month plan.

My company has Github Business, which provides all major models. Claude runs out of tokens usually in the middle of the month. I can use it with opencode.

Then one extra Claude Pro. Then also:

Codex business for a test month. gpt-5.4 was excellent.

And since kimi-2.6 I bought their monthly for testing. Works fine. Not as excellent as opus, but usable. Fixes tricky problems by its own.

Not an OpenRouter subscription yet. Those models are either free on opencode anyway, or not good enough. The free MinMax 2.5 on opencode is pretty usable also.

I just cancelled before seeing this news. i was already pissed about constantly hitting limits on the 20 a month plan and looking for alternatives and this seals the deal. Bye bye!

I just paid for Pro for the first time 24 hours ago. Its been great, but the limits are crazy. It's nice not dealing with ChatGPTs sycophantic gaslighting, and not having random bugs.

That said, I seem to be caught in that 2% test if I open in a private tab. What nonsense. I wouldn't be paying for Claude if it wasn't for its quality abilities, which necessarily includes Claude Code.

I can easily hit the weekly limit on Claude even on the $200 plan. I have yet to ever hit a rate limit on Codex $100. And the results are almost as good. And don't get me started on Anthropic's extra usage scam.

How do you hit that limit? I’m never close to it.

Not the op, but it’s fairly easy to hit if you automate a kanban and have some stuff you want to get done. All those little “wouldn’t it be great if” tasks that show up after doing a big task become very doable, it just soaks your tokens.

With Opus 4.6 on the $20 plan the limits were bad, but at least you could do a short session.

I find that with Opus 4.7 I can do two messages. Once I had a short session with 4-5 messages and it consumed $10 in extra usage.

This relegated Claude to a backup option in addition to Codex, which has the better desktop app anyway, and much better usage limits.

I’m considering to even cancel Claude entirely.

Yea, I've been fine so far, but something happened with Opus 4.6 and especially 4.7. I was able to do some actual work with a Pro plan before. Now it's just pure anxiety of hitting the limits.

With Sonnet it's a bit better, but I can get the same performance with GPT-5.4.

Now I'm pretty much paying the 20€ for Claude Pro so it can plan/review stuff and then I use pi.dev + GPT-5.4 for the actual work.

Reply on news.ycombinator.com