I don't marginal intelligence has diminishing value -- quite the opposite actually. Every additional "IQ point" or 9-of-reliability unlocks some new economically valuable task. But in a few years we'll know whether or not that's correct!

This has echoes of the same argument that why hire anybody in the US when you can hire someone almost as good elsewhere for a small fraction of the cost.

In winner-takes-all markets, it becomes worth it to pay 10x for 10% better.

Companies indeed hired someone (Asian) almost as good elsewhere for a fraction of cost.

> This has echoes of the same argument that why hire anybody in the US when you can hire someone almost as good elsewhere for a small fraction of the cost.

The reason in workplaces people don't hire somebody else at a fraction of the cost isn't because American workers are 15% smarter than non-Americans. It was just very difficult to have exactly 1 remote worker on a team, language barrier, international laws, etc.

In my experience people generally aren't willing to pay double the salary for 15% better work. They will however pay market for domains where nobody else is qualified.

>people generally aren't willing to pay double the salary for 15% better work

Actually, companies are already willing to pay double for 15% better work - the proof is in what SWE salaries in America have currently become.

Companies are pretty much paying at the frontier of marginal improvement (not returns, which again, are hockey-stick) - i.e. it's not like paying 800k will reliably hire someone 15% better than the 400k that they're currently paying.

At this point, the limiting factor is their reach-out and interviewing processes, and the ability of the brand name to attract talent (OpenAI can actually get better talent by paying 800k; Microsoft would not because they're still Microsoft and some people would simply never respond to a Microsoft recruiter even if they lead with "this role pays 800k").

Is the winner-takes-all market in the room right now? everybody switching models and harnesses non stop.

Then explain why big corps like MS or Google are outsourcing to development centers in a much cheaper country (India, China).

The explanation is simply that it's a mistake.

American software has flip-flopped between rushing to outsource and realizing that outsourcing is a bad idea. LLMs (plus US work visa policy headwinds) are simply what triggered the current flip cycle.

> I think it will end up bad for them

There will always be a market for the most intelligent model. Keep in mind as well that something like Fable doesn't have to be cost effective on its own. Fable is Anthropic's way of building datasets for smaller more dense models, which is viable on its own as a method of success. These large labs are constrained by their dataset more than anything else, and these expensive-but-intelligent models create some of the best data out there.

I agree after using DeepSeek for many tens of billions of tokens for this point. Yeah Fable/Astra has more "magic", but not running out of token budget 5 days into the month is the right place to be. Plus, learning to deal with dumber models, only keeps you sharper as the programmer/agent wranger.

Opus 5.5 still exists

As per OpenRouter Opus 5.5 costs $4 / M tokens and DeepSeek Flash is $0.006 / M tokens, which makes it x666 times cheaper (I urge you to re-check my math!)

https://openrouter.ai/compare/anthropic/claude-opus-5.5/~dee...

To be honest, I use both daily and feel no difference in this case

> few percent better

A few percent better on tests like this translates to extreme real-world differences, and that's not even taking into account how benchmaxxed DeepSeek models tend to be.

Using DS-Flash is like using local models that use all your hardware.

It's fine for slow tasks but if you need something done quickly, I find that DS just spents at lot of time thinking.

After using Claude for the first time for about 2 weeks, I've been impressed by the quality and speed of Sonnet 5.5 to do code changes (~20 seconds). Meanwhile DS would have taken 2-5 minutes.

You can get ds flash providers that do 1000+ toks/sec - it’s unreal fast.

The distilled ones which are even worse than the original?

You will get more done, at higher quality and quicker with an Anthropic sub compared to using DeepSeek. Unless you’re forced to use API billing the open weight models make no sense right now.

A lot of companies ARE forced to use API billing though, as Anthropic only allows up to 150 users to use a sub. And companies are Anthropics biggest revenue stream - criticizing the cost per task of their models is a valid criticism.

This feels like saying you'll be more successful if you drive a Bently to work.

That makes no sense.