Hacker News

points by sigmar 3 days ago | 0 comments

People have been saying the openweight models would collapse anthropic's revenue since the deepseek moment in January of 2025. They've stayed ahead in model intelligence and increased revenue by a lot since then.

Not saying it couldn't happen in the future, but we've got to be circumspect about that narrative after almost 2 years of it, no?

[–] afavour 3 days ago

January 2025 isn’t even two years ago yet. When you’re talking about giant companies going public you want to be looking at significantly longer timescales than that.

[–] riknos314 3 days ago

How much of that revenue is usage from hosting partners (AMZN, M$, etc) who have a vested interest in using Anthropic instead of open weights though?

[–] JacobAsmuth 3 days ago

Well then the fact that those hosting partners have 10+ year commitments should make you pretty confident then, right?

[–] cruffle_duffle 3 days ago

The thing holding open models back is the insane cost of gpus. If/when that cost goes down… look out closed models!

[–] grumple 3 days ago

In January 2025 most of us didn’t even consider the frontier models good enough for regular usage.

[–] larksimian 3 days ago

Claude max sub is like 4x cheaper than deepseek equivalent, isn't it? At least 2x cheaper.

[–] ex-aws-dude 3 days ago

Why would you be circumspect if the gap kept getting closer?

Doesn't that indicate its becoming _more_ true?

[–] sigmar 3 days ago

What is the evidence the gap is closing? Seems like it has been the same "6 months behind" since Jan 2025. The first fable was trained in march and we still don't have an openweight model as capable

[–] bel8 3 days ago

Not everything needs SOTA (State of the Art).

If you control for cost, not only the gap closed, but it switched already.

DeepSeek 4.1 Flash dominates perf/cost for suitable tasks.

And if you do need SOTA, Kimi 3.1 Pro scores 55 on https://artificialanalysis.ai vs 58 of Opus 5.5. And they were released 1 month appart.

Kimi even outperforms Opus on some tasks because it doesn't downgrade its intelligente on cyber and cryptography workloads or ban you if you use it for LLM research.

So the gap is 1 month, or zero, depending on your needs.

[–] chasd00 3 days ago

> Not everything needs SOTA (State of the Art).

but if they cost the same then who cares if it's SOTA or not? If I can use fable to solve a problem a lesser model is capable of but i have fable at my finger tips at about the same cost then why would i go through the effort to switch?

[–] bel8 3 days ago

Since when DeepSeek and Fable cost the same?

Even Kimi K3.1 costs half of Opus 5.5 on openrouter which is the closest indication of cost that we have.

[–] kroaton 2 days ago

Speed.