Ironically the few people not scamming you for cache reads are Deepseek.
Everyone else charges a ridiculous amount but Deepseeks API is $0.003625 / M tok.
I'm surprised no one talks about this because of how significant it is. GPT 5.5 for example costs a ridiculous $0.50 / M tok cached. It's literally almost 140 times cheaper which matters a lot for tool calls.
doesn't matter when subscriptions get cache reads for free, it is only really worth it if it's x340 cheaper otherwise I'd be paying $120 a day, 90% of the cost being cache reads for any top level opensource model.
The details are secret. It very well could be wasted GPU time but Anthropic could have made a killer offering as well.
I'm just speculating, but a particularly killer offering Elon wouldnt be able to refuse would be if Anthropic agreed to give them some training data / technology.
> American Efficient then used that sales data to calculate the energy savings from the anticipated use of the lighting and appliances, entering those projected savings into “capacity auctions.”
> At capacity auctions, grid operators pay for the ability of traditional power suppliers and utilities— as well as energy-efficiency aggregators like American Efficient—to produce power when needed.
The home depot example shows it more succinctly. How does American Efficient sending a small check to home depot mean that they get to bid for having produced capacity?
If I squint I can almost imagine the goal of this setup. If you want people to use less power you could definitely promote energy efficient appliances and lighting via market forces.
But doing so at capacity auctions seems ridiculous. If your power company wants it then they should cut checks directly to the consumer as a discount/subsidy on energy efficient appliances.
10M tokens of raw execution traces to grep through is slop. The tasks are fizzbuzz, palindrome, list reversal, and sum-even. The palindrome challenge is literaly this:
> Is the word "racecar" a palindrome? Answer with exactly one lowercase word: "yes" or "no". Print only the answer.
It would also imply that it costs Google ~7¢ in only energy cost to deliver that file to you (using average EU energy costs), which is clearly non-sensical given the rates hyperscalers charge for network egress.
Additionally, the cited number also conflates wired internet (low power consumption) with mobile internet (higher), even though this model is only being downloaded to Chrome Desktop AFAICT.
When looking at the power consumption across the whole network path and not just a single link, most of the power draw is probably baseline static power costs of keeping all the routers and switches running. Which means that judging the impact of a download in terms of Watts per MB/s is a pretty bad way of analyzing this.
> Showing 1,808 changed files with 790,916 additions and 151 deletions.
Just looking at the git diff [0].
I looked at one of these rust port files [1]. Its 827 loc and apparently 7,576 tokens. So that gives you a first order guess that the full 700k additions is around 8 million output tokens. Obviously there are some tool calls, reasoning, reads of the zig version, and fixing compile errors as overhead. So I would guess maybe this is like 40 million tokens by multiplying by 5?
If we guess that is around $200 to $500 in token spend. We can probably guess that it emits around the same as buying $100 in gas? Or like 50 or so kgs of CO2?
Income through taxes is roughly a percentage of GDP.
You could also just compare interest spending vs budget, and lots of people do. Spending on interest is roughly $1T out of a total $7T with income of $5.23T
$1T of an incoming $5.23T is pretty concerning. Especially given projections that the $1T is likely to go up significantly over the next decade.
It's not hard to imagine how this happens. I assume most people here have used these models extensively.
The help bot system prompt probably includes some statement about how Claude should phrase everything as "we".
The system prompt includes statements about how it doesn't have tools for managing funds.
A little bit of A and a bit of B and you get a message from Haiku telling you that you can't get your money back said as though this isn't a trivial customer service thing to do.
> The help bot system prompt probably includes some statement about how Claude should phrase everything as "we".
Yes, why did Anthropic do that when everyone knew it could result in this situation we're discussing?
> The system prompt includes statements about how it doesn't have tools for managing funds.
Yes, why did Anthropic do that when everyone knew it could result in this situation we're discussing?
What you've been describing are all effects of the cause, which is poor management decisions to have poor support and poor customer service. Clearly those decisions resulted in poor support bot system prompts, too.
To wit: this would likely not have happened if the prompt included something like "in a scenario like this, or any scenario where the customer asks, simply transfer them to a human", and if Anthropic had not decided to have dysfunctional support and customer service.
The feedback from folks here is not that poor decisions can have poor effects. It's 'for the love of god, please stop making poor decisions that repeatedly, invariably, lead to unforced errors like the one in TFA'.
On an electric car that yells at you your range left and that you won't make it to your destination unless you charge, if you turn on the seat warmers, that range goes down so you have to think about if you'd rather have a toasty butt and have to stop and charge, or just be colder and get there sooner. But you have to charge anyway.
In an ICE powered car, running the heater doesn't have the same effect on range. Because an ICE is hot due to how it works, sending hot air to the car's interior is basically free because the heater uses waste heat from the engine.
Everyone else charges a ridiculous amount but Deepseeks API is $0.003625 / M tok.
I'm surprised no one talks about this because of how significant it is. GPT 5.5 for example costs a ridiculous $0.50 / M tok cached. It's literally almost 140 times cheaper which matters a lot for tool calls.
reply