Haiku ou GPT Luna: qual funciona melhor no Hermes?

Author: Cortes do Rafael Quintanilha · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Haiku ou GPT Luna: qual funciona melhor no Hermes? · Channel: Cortes do Rafael Quintanilha · 07:45 · Watch on YouTube

Topics: AI agents, AI developer tools, AI models

Summary

The speaker explains how to connect Hermes, a personal assistant, to Anthropic models using an API key and credits. He argues that API credits avoid using a Claude subscription’s normal limits, while direct subscription options can require extra credits or unreliable workarounds. He demonstrates configuring Claude Haiku 3.5 in Hermes and testing it with YouTube-related queries, noting its speed but also its lower proactivity than larger models. Comparing it with GPT-4o Luna, he finds Haiku 3.5 much faster and changes his recommendation for using Hermes.

Key points

Hermes can be configured to use Anthropic models through an API key.
Using API credits keeps Hermes usage separate from the normal Claude subscription limit.
The Claude subscription route in Hermes may draw from extra credits, and a subscription plugin can require ongoing maintenance.
The demonstration configures an API key and enters the custom model name claude-3-5-haiku.
The speaker reports that Haiku 3.5 answered a YouTube query at 118 tokens per second, then another at 136 tokens per second.
Haiku 3.5 needed a hint to identify the correct older Jev video, showing less proactivity than models such as Opus and Sonnet.
In the speaker’s comparison, GPT-4o Luna generated at about 13–15 tokens per second.
The speaker recommends Haiku 3.5 for Hermes users seeking a fast personal assistant.

Chapters

Start	Chapter	Summary
00:00	Using API credits with Hermes	The speaker introduces using Anthropic API credits in Hermes and logs into the VPS where it is installed.
01:01	Subscription limits and API credits	He explains the available Claude setup options and why API usage can be preferable to relying on a subscription.
03:04	Configuring the Anthropic API key	The speaker describes subscription workarounds, creates an API key, and configures Haiku 3.5 as a custom model in Hermes.
05:09	Testing Haiku 3.5	He asks Hermes about videos on his YouTube channel, demonstrating Haiku’s speed and its need for a hint on one query.
06:43	Comparing Haiku 3.5 with GPT-4o Luna	The speaker compares response speeds and concludes that Haiku 3.5 is the better choice for his Hermes personal assistant use case.

Quotes

The first advantage is that, of course, you won’t be tapping into your Claude subscription limit — 01:31

So you have to enter the custom name — 04:37

it lacks some of the proactivity you’ll see in models like Opus and Sonnet. — 06:12

you should use Haiku 3.5, because it will be very fast — 06:43

We are talking about Haiku being 10 times faster than Luna. — 07:14

Concepts and entities

Hermes, Anthropic, Claude Haiku 3.5, GPT-4o Luna, HostGator, Jev, API key

Full transcript (original language)

00:00 The first thing I wanted to show you is how you can use these API credits in a personal assistant, like Hermes, okay? Why? There has traditionally been some confusion as to whether or not you can use your Claude subscription in third-party projects and apps, right? Like in the case of Hermes. OpenAI has always been quite open about this and has allowed you to use your subscription in these other apps. Anthropic has always had a more hostile

00:31 attitude toward this. But now, since they are offering you this value in the API, you can configure your account using an API Key and use it in other services, like I said, Hermes, okay? So , what are we going to do here, right? I’m going to log into my HostGator VPS, which is where my Hermes is installed. My recommendation for anyone who wants to have a personal assistant running 24/7 is to get a VPS and install it there. If you are going to choose a VPS , I recommend that you do it at HostGator, which is our partner here on

01:01 the channel, okay? The link for you to hire your VPS with 24/7 support in Portuguese will be in the description and in the pinned comment. So, I’ve logged into my VPS and, well, I can initialize Hermes here, okay? And you will see that I now have Claude 3.5 Sonnet connected. How did I do this, guys? I’ll exit and show you that I can run this: Hermes model, and I have the Anthropic option selected, see? Claude Models via API Key or Claude

01:31 subscription. So, if I select this, it will ask if I want to configure it with my Claude Pro or Max subscription, or if I want to use an API Key. Wait, Rafael, if I can use my Claude Max subscription, what is the advantage? Why would I use an API key? Well, the first advantage is that, of course, you won’t be tapping into your Claude subscription limit, right? This is often important because you don’t want to mix one thing with another, or you’re sharing an account, a personal

02:01 assistant, with your company or another person. So, having an API key allows you to do that. But that’s not even the most important thing, okay? Why? Because even if you manage to configure it there with, what ends up happening, okay? When you select Claude in Hermes, and I’m reading this to you from the official Hermes documentation, what happens? Hermes does this routing, like Claude on your account, but it only works if you’re on a Max plan and have

02:33 purchased extra credits; in other words , it won’t draw from your normal limit. It will draw from your extra credits. What is that? If I go here to my Anthropic account, you’ll see that I have here, look, in my section, user credits that I can use as extra credits , okay? So I can buy credits here, see, “buy more”, I can add fuel here inside my Anthropic account and then yes, I would be using this subscription of

03:04 mine, but drawing from here, okay? It wouldn’t be drawing from my normal limit. So this doesn’t work for most people. There is another way to do it with a Hermes plugin, okay? So there is a Hermes plugin here called Hermes Plugin Claude Subscription Direct, which is also managed by the Hermes team themselves. But again, folks, it’s that situation where you’re always living in limbo. Will it work? Won’t it work? Then it stops working,

03:35 you have to keep updating it, and it’s a pain, okay? So I never liked using Anthropic models in Hermes because I couldn’t use my subscription, and I was also tired of having to rig things to make them work. Thank God, now that we have the API Key with credits, I can go to my Claude account. So I can come here to platform.claude.com. I come here to API keys, I generate an API key . Look, you can see that I even have credit, I have 224 in credit. And look,

04:06 I’ve been testing all day, I haven’t spent a single dollar, okay? See, I spent 96 cents, right, in all my tests. So, I can generate my API key here, go back to Hermes, and put it in there. Got it? So, by doing this, I configure Hermes. I’ll even do it with you here because there’s a detail, okay? If you select the API key option here, it will ask you to enter it, right? So I’m in the workspace here, I’ll create a key, continue with API here, I’ll name it hermes, and I’ll select that I don’t want it to expire, right? Never expire. Alright, that works. I’ll create the

04:37 key. I copied the key here. You didn’t see it because you shouldn’t share your API key. I pasted it and it’s great. Look at that. But there is a detail, look. If you look here, Hermes hasn’t updated yet, so Haiku 3.5 won’t show up . See? It’s only showing Claude Haiku 3.0. So you have to enter the custom name, okay? It might be that when you’re watching this video, Claude Haiku 3.5 hasn’t appeared yet. So you come here to enter custom model and put claude-3-5-haiku. Got it? Select high

05:09 here. Wonderful, it’s already running. So I can initialize Hermes now and I can ask something. For example, what was the latest video on my YouTube channel? Got it? It has access to a tool I have that controls my YouTube. Oh, it already grabbed it, it was able to load the skill, it made the call. There it is, the video I just published . And look at this, folks, 118 tokens per second, okay? It took basically 8 seconds to give me this answer, okay? I can ask something else here, look. How

05:41 many views did the Jev video, the first one, have? This was the most viewed video on my channel so far. So I’m asking it to do another query there. Eh , no, it’s grabbing, it grabbed the wrong one, right? Not this one, the other one. So, let’s see if it will be able to understand this. Not if you need to search. I think it was within the last thirty days. So here you notice, right, that it lacks some of the proactivity you’ll see in models

06:12 like Opus and Sonnet. It’s not that smart, but there you go, finally, after a little hint, it managed to get it, which one? The old one, which is the first one from Jev. And there it told me the number of views. Got it? What I wanted to show you here is to do several interactions so you can see the speed. So, look at that, 136 tokens per second. Man, very good. So, I am officially changing my recommendation, if it’s worth anything. For those of you who want to use a personal assistant like Hermes, you should use

06:43 Haiku 3.5, because it will be very fast , much faster, for example, than Luna, okay? So, if I kill this session here, look, and initialize Hermes again, and then we’ll change the model to ChatGPT with GPT-4o Luna. Let’s set it to Haiku , right, which is the same. And then we’ll ask the same question, okay? Uh, what was the latest video on my YouTube ? Right? It was that first question I asked. And then we’re going to compare the speed here. Look at this, guys, 13

07:14 tokens per second. It makes you want to cry, doesn’t it? 15 tokens per second. We are talking about Haiku being 10 times faster than Luna. So, it’s a huge difference when you’re using a personal assistant. Look at that, very, very, very slow, okay? So, guys, look at this . 15 tokens per second. It’s not even returning the answer yet. I’m actually irritated by this. I’m going to stop the test. What I wanted to show you is the difference in speed. So, I always thought Luna was the best model to use with a personal assistant like Hermes.

07:45 I don’t think so anymore, okay? At the time I’m recording this video, if you have a Claude Pro subscription, if you pay more than $ 20 for your subscription, it’s certainly worth using a personal assistant like Hermes using those API credits with Haiku 3.5. Yeah.