Not only is Anthropic (and other AI providers) losing money on
- Competing with other AI firms
- Selling their own product
- Buying hardware with ever increasing costs
They now have another problem, which is the open source models, or as the popular parlance now calls it "The Chinese models". DeepSeek v4 has been released for a while, and it is surprisingly good for its price:
| DeepSeek V4 pro | Sonnet 4.6 | Opus 4.7 | |
|---|---|---|---|
| 1 million input tokens | $0.43 | $3 | $5 |
| 1 million output tokens | $0.87 | $15 | $25 |
Prices are without taking token caching into account, and I didn't include Haiku and Flash, as I don't consider them good enough for coding. YMMV. DeepSeek did have a 75% discount for a promotion period, but then they decided to make that the permanent price.
DeepSeek is not yet up to something like Opus 4.5, but it comes very close. I know Opus is on 4.7 now, but 4.5 was the first one where I admitted that "ok, this is actually working, AI coding is here to stay". Everything after that felt like just tiny improvement for a huge price increase to me (due to the increased token usage).
This is a massive price difference. Sure, it's not the best, but unless you are oneshotting problems with no human in the loop, this is an insanely good price and well worth exploring. I would also question how much of a con is the human requirement really is. I think at this point, any coding model will produce something that seemingly works on the first glance, but subtle bugs, security vulnerabilities, or edge cases would still be present. I wouldn't put anything that any model output's on the internet without reviewing the code first, this is not specific to DeepSeek.
The models are open, available for everyone to run. While it's still not possible to run coding models (as far as I know) on consumer hardware, you can run DeepSeek, QWEN coder, and other models on the cloud. It is impossible to know where we are in the curve (just starting an L curve, or are we at the peak already?) but if things keep advancing at this rate, running very capable coding models on Macs, or Framework desktops will arrive sooner than I expected. A local model with 0 telemetry or spying that you can tweak in any way you want. That is the hacker dream.
Meanwhile, Anthropic is tightening the screws harder and harder, people are reporting hitting limits sooner and sooner, degrading model outputs, token usage going up with every tiny model improvement, Microsoft admitting that they cannot keep up with increasing AI costs, Uber burns through their yearly AI budget in just a few months (see this), people running their own LLMs in the cloud, and eventually on their local PCs, what next?
Anthropic even tried removing Claude Code from the pro plan at some point, and claimed it was just a prank bro an A/B test. They clearly are looking for ways to at least slow the bleeding.
There is also the non negligible moralistic angle. Every AI firm is basically profiting off from all the free work available on the internet (or in Meta's case, they just bittorrent all the books). It's a really good tool to strip those pesky GPL licenses from projects. I feel less bad about letting DeepSeek slurping up my code, since at least some of that gets returned to the community in the form that the models are open and available for everyone.
So what is next for Anthropic and the rest of big tech who bet so heavily on AI?
As a sidenote, if you are interested in a more detailed look into the economics of AI firms and how infeasible their burn rate is, I can really recommend Ed Zitron's blog.