Hacker Newsnew | past | comments | ask | show | jobs | submit | minimaxir's commentslogin

I questioned OP's "there is an answer online" claim so I checked and the only source found for the original question was a 5th grade Russian school for mathematics.

https://aistudio.google.com/app/prompts?state=%7B%22ids%22:%...


If your math does not involve multiplying 20 digit numbers, modern LLMs can "do" math even without a Python tool despite the counterintuition of next token prediction.

And if you give your LLM access to a calculator, it will have to problem multiplying 20-digit numbers.

Hugging Face is most definitely not bootstrapped.

There's more going on here than just the "routing" part.

That's why I gave it a month and not an afternoon.

Credit is prepaid at a 5% surcharge.

OpenRouter has stealth models that are indicated as "by OpenRouter" but indicate an external provider.

https://openrouter.ai/openrouter/owl-alpha


As someone who uses OpenRouter extensively (and wrote an unintentional adjacent PR piece a few days ago: https://news.ycombinator.com/item?id=48317294 ), it's definitely the best way to try out new models without fiddling with each providers distinct APIs which is becoming a recurring concern as of late.

That said, I don't understand the people who use something a full agentic backbone with expensive models like Claude Opus with OpenRouter because that 5% surcharge is meaningful at that level of cost instead of going with the source API providers. But people are clearly doing it, and it's pure revenue.


IDK, but that sounds like something that would be better implemented with an open-source library to which providers supply support patches. Why do I need a company to act as a proxy and not just run a relatively simple shim layer on my machine?

I'm just a stupid systems programmer working in the bowels of AI and I understand there is a lot of seemingly pointless software which exists solely to provide a slight boost to convenience in exchange for money. Is OpenRouter just that? Do they actually host models themselves or centralize billing amongst various providers?


A library with a bunch of different providers doesn’t solve the payment/billing problem (which is one of the main openrouter benefits). IMO being able to buy credits and not have them locked to one provider is worth the 5% to me.

There is a lot of dumb token spend right now - tokenmaxing and such. Economic cost of token is not being evaluated carefully because there is fomo and no one wants to be left behind. But folks are waking up to it, and dumb token spending is not sustainable and will revert.

Better uptime? Given it will be routed to one of Anthropic, Amazon Bedrock, Claude Platform on AWS, Google Vertex (Europe) or Google Vertex

That seems also very possible to do in a client library. (Though that would benefit from an across-my-nodes gossip of backend statuses.)

im happy to pay 5% extra for consolidated billing and usage limits. It just makes it easier

Convenience has a markup

>> it's definitely the best way to try out new models without fiddling with each providers distinct APIs which is becoming a recurring concern as of late

Why not... Cursor?


Cursor only supports a single model (Kimi K2.5) not made by the Big 4 labs (OpenAI, Anthropic, Google, xAI). Cursor is actually extremely bad at wide model support.

OpenCode is much better at it.


And its own model (Composer 2.5)

Which is finetuned Kimi 2.5

Cursor is functionally "one of the big 4 labs" (SpaceX).

First, calling xAI one of the big labs is pretty funny.

Second, Cursor hasn’t been acquired by SpaceX yet, and there’s a good chance they never will be.


I use Cursor with OpenRouter for some projects and it's great. Most of the time I just use Auto and let Cursor use its model or choose. If I run out of quota, or I'm not getting what I want, I switch off Auto and use OpenRouter to pick Opus, Codex, or whoever(all are available). Can continue the same context if you want, type "please continue" in the agent prompt, and on you go.

Cursor has limits even when using your own key. I was even cut off using a local model. I guess they use some sort of harness that requires non-local resources? I'm not sure I've actually tried to use Cursor in a fully-offline scenario yet. Cline works well enough and doesn't require any sign-up.

Cursor's coverage on open weight models is very minimal, and it's irrelevant for testing models in your actual application.

The post goes into that issue. Throughly.

The numbers at the beginning of the post are weekly aggregate values well after the endpoint was paid-only.


The post is wrong, it's still free, see - https://openrouter.ai/tencent/hy3-preview:free it's free in kilo.ai https://kilo.ai/models/tencent-hy3-preview-free It's free in a lot of places.

The first endpoint was closed. If you actually try and call it from the API you get this response:

> Hy3 preview is no longer available as a free model. It has transitioned to a paid model. Continue using it here: https://openrouter.ai/tencent/hy3-preview

The Kilo Code may have free traffic but if you check the numbers is still inconsequential relative to the trillions of tokens through OpenRouter.


One idea I had was to count # of distinct API keys that have spent atleast $100 (number's flexible), which would be enough to provide guidance on if the traffic is from a single power-user.

In the Cursor case which is BYOK, that would count as distinct API keys.


Deception is not ideal for agentic coding.

Yet if parent is right, the capacity to deceive might be a strong heuristic for the things you do care about.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: