Completely agree on your first point: software development is so much more than writing code. LLMs are a threat to programmers for whom the job is 8 hours a day of writing code to detailed specifications provided by other people. I can't remember any point in my own career where I worked with people who got to do that.
There's a great example of that in the linked post itself:
> Let's build a property-based testing suite. It should create Java classes at random using the entire range of available Java features. These random classes should be checked to see whether they produce valid parse trees, satisfying a variety of invariants.
Knowing what that means is worth $150/hour even if you don't type a single line of code to implement it yourself!
And to be fair, the author makes that point themselves later on:
> Agentic AI means that anything you know to code can be coded very rapidly. Read that sentence carefully. If you know just what code needs to be created to solve an issue you want, the angels will grant you that code at the cost of a prompt or two. The trouble comes in that most people don't know what code needs to be created to solve their problem, for any but the most trivial problems.
On your second point: I wouldn't recommend betting against costs continuing to fall. The cost reduction trend has been reliable over the past three years.
In 2022 the best available models was GPT-3 text-davinci-003 at $60/million input tokens.
GPT-5 today is $1.25/million input tokens - 48x cheaper for a massively more capable model.
... and we already know it can be even cheaper. Kimi K2 came out two weeks ago benchmarking close to (possibly even above) GPT-5 and can be run at an even lower cost.
I'm willing to bet there are still significantly more optimizations to be discovered, and prices will continue to drop - at least on a per-token basis.
We're beginning to find more expensive ways to use the models though. Coding Agents like Claude Code and Codex CLI can churn through tokens.
I get your point, but I don't think the pricing is long term viable. We're in the burn everything to the ground to earn market share phase. Once things start to stabilize and there is no more user growth, they'll start putting the screws to the users.
I said the same thing about Netflix in 2015 and Gamepass in 2020. It might have taken a while but eventually it happened. And they're gonna have to raise prices higher and faster at some point.
I think it's gonna be a lot worse with LLMs. Mainly because they're substantially under charging, I don't think the cost of operating is going to drop much, and the workflows are only going to get more token hungry.
The incentives here are also fucking atrocious. They aren't incentivised to make the model as good as possible. It's in their best interest to tune so it's good enough to not drive you off, but bad enough to push you to spend more.
It wont: if/when LLMs start to get too expensive, people will just migrate to open models, run it local, etc. I see no scenario where we are held hostage by the main providers.
I was a lot more worried about this when only OpenAI and Anthropic had truly great models - but now we also have Google and five different Chinese AI labs who are releasing open weight models that are in the same ballpark as the OpenAI and Anthropic ones. I think we'll be fine.
Very much agreed - right now someone might hold a few months of competitive lead, but open models catch up fast. Plus the lack of any real vendor lock-in means there's just not room for extortionate pricing.
> In 2022 the best available models was GPT-3 text-davinci-003 at $60/million input tokens.
>GPT-5 today is $1.25/million input tokens - 48x cheaper for a massively more capable model.
Yes - but.
GPT-5 and all the other modern "reasoning models" and tools burn through way more tokens to answer the same prompts.
As you said:
> We're beginning to find more expensive ways to use the models though. Coding Agents like Claude Code and Codex CLI can churn through tokens.
Right now, it feels that "frontier models" costs to use are staying the same as they've been for the entire ~5 year history of the current LLM/AI industry. But older models these days are comparably effectively free.
I'm wondering when/if there'll be a asymptotic flattening, where new frontier models are insignificantly better that older ones, and running some model off Huggingface on a reasonably specced up Mac Mini or gaming PC will provide AI coding assistance at basically electricity and hardware depreciation prices?
That really is the most interesting question for me: when will it be possible to run a model that is good enough to drive Claude Code or Codex CLI on consumer hardware?
Also, at some point the Blackwell-generation DGX Station is supposed to ship with 768 GB of unified memory. It will presumably come with a high five-figure price tag, and it should be able to run most open-source models with little need to trade off quality for speed.
Trouble is, there's not even much hype surrounding the launch yet, much less shipping hardware. Which seems kind of ominous.
> Knowing what that means is worth $150/hour even if you don't type a single line of code to implement it yourself!
Yeah but eventually there wont be enough people who actually do know all that.
The hope amongst the proponents is that by the time that happens they wont need anyone who knows that because SOTA will have replaced those people too.
There's a great example of that in the linked post itself:
> Let's build a property-based testing suite. It should create Java classes at random using the entire range of available Java features. These random classes should be checked to see whether they produce valid parse trees, satisfying a variety of invariants.
Knowing what that means is worth $150/hour even if you don't type a single line of code to implement it yourself!
And to be fair, the author makes that point themselves later on:
> Agentic AI means that anything you know to code can be coded very rapidly. Read that sentence carefully. If you know just what code needs to be created to solve an issue you want, the angels will grant you that code at the cost of a prompt or two. The trouble comes in that most people don't know what code needs to be created to solve their problem, for any but the most trivial problems.
On your second point: I wouldn't recommend betting against costs continuing to fall. The cost reduction trend has been reliable over the past three years.
In 2022 the best available models was GPT-3 text-davinci-003 at $60/million input tokens.
GPT-5 today is $1.25/million input tokens - 48x cheaper for a massively more capable model.
... and we already know it can be even cheaper. Kimi K2 came out two weeks ago benchmarking close to (possibly even above) GPT-5 and can be run at an even lower cost.
I'm willing to bet there are still significantly more optimizations to be discovered, and prices will continue to drop - at least on a per-token basis.
We're beginning to find more expensive ways to use the models though. Coding Agents like Claude Code and Codex CLI can churn through tokens.