A OpenAI acabou de matar o Jev? (Decisions API)

Author: Matheus Battisti - Hora de Codar · 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: A OpenAI acabou de matar o Jev? (Decisions API) · Channel: Matheus Battisti - Hora de Codar · 24:23 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI developer tools, Multimodal models

Summary

The video introduces OpenAI’s Decisions API, explaining how decision-focused AI selects from predefined answers and how it differs from a conventional LLM. The presenter compares it with Jev on price, speed, and capabilities, then integrates the API into a veterinary-triage app using true/false, choice, and score questions. The demo also adds image analysis and shows the returned decisions, confidence levels, latency, and estimated cost. The presenter concludes that the API works well and supports images, but seemed slower and more expensive than Jev in these tests; a direct quality comparison remains open.

Key points

A decision API returns a choice or score directly, rather than a long natural-language response.
The API supports true/false, choice, and score question types, and can analyze text, images, or both.
The presenter describes Luna as the model available for the Decisions API.
The comparison presented says Jev is cheaper, while OpenAI’s API is slightly faster in principle; the speed difference was described as practically imperceptible.
The integration uses an OpenAI API key and can be done in any system that can call the API; it is not tied to Codex.
The demo integrates three decision questions into a veterinary clinic triage dashboard and displays confidence, response time, and estimated cost.
Answers with confidence below 60% are highlighted for review in the demo.
For image messages, the demo sends the photo with the text and adds a question about visible injuries, bleeding, or swelling.
The presenter found that image analysis worked and cost only slightly more in the demonstrated example.
The video does not establish which service has better answer quality; the presenter invites a future head-to-head comparison.

Chapters

Start	Chapter	Summary
00:00	What the Decisions API does	The presenter introduces the API and explains decision AI, its response types, pricing, and how it differs from a conventional LLM.
05:41	Comparison with Jev and documentation	The video compares the services’ stated speed and costs, then covers the public beta, documentation, API key, and playground.
10:53	Integrating the API into a triage app	The presenter builds a veterinary triage workflow with three decision questions, confidence-based review, and cost and latency tracking.
15:33	Testing triage results	The app is tested on pet-care messages, with the presenter reviewing the returned classifications, confidence, and response times.
17:07	Adding image analysis	The presenter extends the app to accept photos, analyzes them alongside message text, and checks the results and token costs.
21:17	Wrap-up and overall assessment	The presenter recaps the documentation and demo, describing the API as simple to integrate, slightly slower and more expensive than Jev in the tests, but notable for image support.

Quotes

An API, for those not in the technical field, is basically software that we can integrate with, send requests to, and receive responses from. — 00:31

We can send text, image, or both together for analysis. — 03:08

we don’t need to use Codex, we can do it in any system, because it’s integrated via API. — 08:44

It costs a bit more than Jeev and it accepts images, right? — 23:53

Concepts and entities

Decisions API, OpenAI, Jev, Decision AI, Claude, Codex, Pet On-Call, GPT-6 Luna

Full transcript (original language)

00:00 Gara released the Decisions API. Now, we can actually use it. You just need an OpenAI API key, the creator of ChatGPT, and you can easily integrate a Jev competitor into your own project, test it, and see if it’s worth it. In this video, I’m going to show you everything about the Decisions API, how much it costs, how to integrate it, and also build a project so we can see in practice if it’s worth it or not. So come on in, let’s go. So guys, let’s first talk a little bit about the API,

00:31 okay? For those who don’t know, an API, for those not in the technical field, is basically software that we can integrate with, send requests to, and receive responses from. That’s basically it. So, that’s why this name might sound strange to some—decisions —because we’re in this era, let’s say , of decision AI, right? As I’ve said in several videos, it won’t replace anything, guys. It adds an extra layer to make our programs cheaper, to actually handle the decision-making part, which was something expensive,

01:03 costly, and time-consuming in LLMs, and consequently, it will make our systems more precise. So we gain another competitor to Jev, right? Several open-source ones have already come out, now this closed one is coming out, which is from OpenAI, the creator of ChatGPT. And it doesn’t give us anything other than decisions, right? Choose between the options we sent, or true or false, okay? So it’s very similar. It has a response time here, right? Approximately 150 ms, similar to

01:34 Jev, right? Apparently faster, but let’s see in practice. A slightly higher price than Jev, at 10 cents per million tokens, but this one is almost infinite. We won’t use many tokens here , okay? And zero on the output, which is the standard for this, because it only delivers one answer, as is the standard for APIs, right? The output works like an API. The decision happens on the input, which is where we pay, right? Well, here is an example of a question. So, if we wanted to know which department this question is intended for or this issue, “I was charged twice on my order,” then we

02:06 have three options, three responsible sectors in the company, and it detects that it’s financial because it was charged twice, okay? So that’s what a decision API does, what a decision AI does. So, check this out, guys, the answer comes out ready without any text in between. So, here is a comparison, right, between a conventional AI, an LLM, and a decision AI. The Decision API here, in this case, could be the JV , the operation is the same. A normal AI, we send a question, a model analyzes it all, reasons, right? We are

02:37 also seeing models reasoning a lot in their answers, okay? And that is good, but for this case here, it is bad. And we get a very complete answer about the question and then the decision, right, all together. So this ends up being time-consuming. The output is always wordy. In decision AI, we send the question with the options, okay? This is standard. Our program is already configured to be like this. A model analyzes here, in the case of the Decision API, it’s GPT6 Luna, okay? It

03:08 is the only one available. And we get an ultra-fast answer with just the selected option and the precision of the answer. This is also important because, like all these models, they can make mistakes, and if the precision is low, we could pass it to a manual review if we wanted to, okay? So as I said, the model that decides in the Decision API is Luna, okay? We can send text, image, or both together for analysis. So it is a differentiator,

03:40 right? In JEV, at least until now, you don’t send images. Questions, what do you want to know and which answers are valid? You have to send that, right? What is the question, right? Like, for example, which department does this person want to talk to? And then the answers: finance, support, and so on, okay? And we receive answers in the return. Like in JEV, we will have three ways to ask, okay? So, the first is true or false? Basically, that is the answer, right? It can only send you

04:10 whether the question you sent is true or false and a precision, obviously, precision is always sent. So, “did it arrive damaged?”, photo of the product, then there is 92%that yes, it arrived damaged, right? Because it will analyze the photo, the question, what we want to do, and then it answers, okay? Choice, then we send what we want to know and the alternatives. So, for example, I was charged twice, which department handles it? The white part is what the customer sent, the gray part is what our system is configured

04:40 for, and the answers as well. So, I was charged twice. The question is, which department handles this? And the answer is finance. A 95%chance of certainty, right? Even though it’s a bit unsure about support, because this could maybe also be solved by support, and so on. And scale, right? They call it score, the name of the question type there changes a bit, right? Predicates there are for JV. Choice is choice and score I think is score as well, okay? And so here it’s as if we wanted a more personalized response, right, where we

05:11 would have to send several options. So we send a scale, right, for example, from bad to good and then from zero to two. So it will fit the response into this scale, to be able to better understand what the person is saying, right? This one is the most customizable possible and can also generate more problems, since the scale isn’t something very fixed, right? And we would have to configure our system to accept various intervals. Well, guys

05:41 , and before we go to the practice here , a brief comparison with JV. In this video here I won’t bring JV along, but if you want, leave it in the comments and I’ll do one with both technologies, okay? So, if we compare the GPT-4o model with Luna, but focused on decision, right? Because Luna also works as an LLM, right? Then JV is a trained model, we don’t know which one it is, but it’s trained for decisions. Question types, true or false? There it is yes or no, but it’s the same thing, right? Choice, choice, and score, okay? Input here. The chart is a bit

06:12 misleading, right? It’s 10 cents per million tokens, which is a lot. And here it’s 0.04. So we have that JV is much cheaper, okay? More than half, right? More than 50%cheaper. So we have that JV is much cheaper, right? It is less than half the price. Speed, as I said, OpenAI is a bit faster. This must be due to the servers, right? It must be delivering more power to at least win on this, because they probably didn’t want to fight on price, okay?

06:42 But this will be practically imperceptible, okay? It’s very little time. That’s the situation. Now we have a public beta, open to any key. Previously, just a few days ago, you couldn’t use it. Now you can. And JV is now available to everyone, right? There used to be a waitlist, now there are no more differences. We can make voice calls through the live API, okay? So this part is cool. There are other more technical issues here that I recommend you read the documentation to understand, like this SDR and the HIPA. Alright, guys? So let’s test the

07:13 Decision API in practice. But just before that, guys, we have our JV course with open enrollment, okay? So, if you want to master these decision tools, as you saw there, it’s all very similar, whether it’s the Decision API or other open-source alternatives, the concept and context are the same. So, by mastering JV, you will consequently know how to operate the others, okay? The link will be in the pinned comment, so take advantage of it while it’s on launch promotion this week, right? Training for you to learn everything about JV, apply it in practical

07:43 projects, and put it into production. Deal? I’m waiting for you there too. Access the page and watch this video here later. Well, guys, so look, OpenAI released this documentation here. It explains how it works, how to integrate it, and what options it has. Here are some practical and visual examples, right? For example, here, oh, this client wants to cancel, please cancel my subscription before the next renewal . 100%probability, okay? So, that’s how

08:13 it works, right? A very practical way. That’s it. Here is the integration format. Here it explains it in another situation, right? Selection of options. Here is the range, right? Which would be the range of values and so on, right ? Well, that being said, if you add a .md to the URL, you get access to this skill, in quotes, okay? So this is the documentation for your agent to learn how to integrate. So remember this, pay

08:44 attention to this detail. Because it will be very important for the integration later, okay? And another thing I want to mention here, we don’t need to use Codex, we can do it in any system, because it’s integrated via API . You will need an API key here on the OpenAI platform, and that is what will enable the integration. It is not tied to the token, okay? Another important thing is that there is also a playground, look. If you access this link here, I’ll try to leave it in the description, folks, okay? It’s also very easy to access, but it has this playground which is just like Jeff’s

09:16 playground. And here you can understand how it works in isolation from your system, right? You simply go there, ah, I want to see how the choice works. So it talks about the customer’s issue, right, the prompt we’re going to send. And here is the choice, right, which would be the mode that gives options. And here are the departments. So, in this case here, we’re asking which team should handle the request, right, the customer’s request there. And it even

09:50 speaks, gives a context, right, choose the team that has the best connection, right, with the main problem, the main issue there of the customer. Then what was it? Oh, I was charged twice for my subscription this month. Can you help me get a refund for the extra charge? So this here is what will be configured in our program on the left, right? And this here is what the AI will receive, right? That the system receives, combines with this, and sends. Then it

10:20 will send it to the Decision API and will receive the response, okay? And we are 100%sure it’s the billing department, alright? So this is what will happen there in our system. And look at this, in a test request on their own server, it took twice as long as they said it would take, right? They even added 200ms for my internet, so their server alone took twice as long as advertised. Start noticing the details there, okay? Alright, let’s

10:53 stop with the small talk and let’s go to the integration. I created a little system here that doesn’t have it yet, okay? I’m going to ask it to start. Let me see here. Ask it to start here, look . It opened the system, folks, okay? And what is this system here? It’s Pet On-Call, a clinic triage. This system has several tickets from a pet shop or a veterinary clinic, okay? And these tickets that are opened in the systems

11:24 will be triaged by the Decisions API, that’s the trick here, okay? So there are 15 in the inbox and they will be analyzed now when I integrate the Decisions API, which is the next prompt I’m going to do. So take a look at how you are going to integrate it. Pay attention to the prompts, okay? I’ll try to leave it in a gist here if I remember, guys; the link will be in the description, okay? Look, now integrate the OpenAI analysis decision API. This first paragraph here, right? You will also implement this first block in your

11:55 system. It is new and you probably don’t know it yet. Before writing any code, read the official documentation. So, this link I gave him is exactly what I showed you here, okay? OpenAI here doesn’t have any explanation, at least I didn’t find it easily, like Jeff, which has that skill for integration, okay? But this should be more than enough. So I pass that link to him and ask him to follow the format there exactly. Use the latest version of the official package, it already has support for it, all right? This here

12:25 teaches him the integration. Then I ask for the functionality. So, upon clicking reanalyze, send the instructor message along with the pet’s name and species. One single call with all three questions together, the yes or no about today’s service that I want to analyze, right? True or false, the subject choice, and the severity rating. So, choice is “choice” and severity rating is the range there, the score, okay? So , I’m going to do the three types of questions in a single request. We can do that. Write the questions and the

12:57 description for each option in Portuguese, explaining well what each option is worth and keep an “other” option for what doesn’t fit. So, here we could perhaps even add manual interaction if it’s “other,” right? So, the call happens on the server. In the browser, show the actual answers on the dashboard, the confidence of each one, which will return that, the time the call took for us to understand, the input tokens so we can see the cost, and the calculated cost. But this is just a rough estimate, because the

13:29 price might change and such, so we’ll see what happens, right? That’s why I included the price in the submission, I’ll show you in a moment. Look, when the confidence of an answer is below 60 , that is something you also have to handle. This is standard even in Jeff. Highlight it in yellow as “to be reviewed.” If the API rejects any question, show that on the screen instead of crashing, right? Just so we can see if there’s any error. Save the analysis to the database and show a badge on the message box with the severity of each analyzed message. Okay then, we requested the prompt, it will handle the integration here. And what do we need to do here in the Env? I

13:59 added the price because it might change , as I told you, just so we keep track of it here. And the OpenAI key. So, how do you get the key? It’s here under API Keys on the platform. Create your key. So, look, I created mine here, okay? And we’ll just paste it into the Env file. So, I’ll paste it here and save it. And then we’ll have it configured. We’ll comment on that later once it returns with the response. Look, I’ve already put it in and saved it. Now I’m just waiting for Claude to get back to me. I’m doing everything in Claude, as I told you. It doesn’t have to be in Codex, right? Just to make this very

14:30 clear, because I know many people will have this doubt. It has nothing to do with Codex. It’s a separate API service , just like creating an image via API. From OpenAI, okay? So let’s wait for it to finish, I’ll be right back. Well, guys, it’s back now, okay? Let’s see what it said. The panel now actually calls the API. The server is running on 3457 and already opening your browser. I tested the call with your key, six messages from the database, and the responses made sense. I couldn’t click on the screen. Not me, anyway. It couldn’t access the browser. Got it. So

15:01 here is what I tested: responses, time and cost, database and badge, key. So, what I’ll do here, I don’t know if it really did it, right? I’ll say here, I put the key in the ENV, start the project here. It said it tested it, but man, there wasn’t even a key there before when it was working, so you can’t really trust it, okay? Let’s just ask it to restart and then I’ll test it myself. So it restarted. Let’s come over here, look. Let’s see, it did one

15:33 analysis, okay? But let’s do the rest, look. I’ll set it to analyze. And here it already returned. So the call time was 300ms, double what’s being advertised, and the cost is negligible, right? So we don’t have that much cost, but even so, it’s double Jeff’s, okay? And let’s see here if it worked. My cat fell from the third-story window, he’s breathing very fast and won’t get up. I’m grabbing an Uber there right now. Needs to be seen today. Yes, 100%sure.

16:04 Great. Emergency is the department, the subject, right? And here, how serious the case seems, it’s an emergency. So everything is fine, look. Let’s see this one here. Good afternoon. Rex’s belly is a bit bloated after eating, and he keeps trying to vomit, but nothing comes out. Must be gas, right? Should I give him gas relief? So it said it needs to be seen today. That took a lot longer. See here? It said it’s an emergency, so I don’t know, right? I think it’s being quite cautious. Let’s try one here that’s a

16:36 bit more…The bird has had its feathers all ruffled all day, huddled at the bottom of the cage. It didn’t even sing today. Could it be cold? Look , needs to be seen today. Emergency, but here, look, for example, 100%was 99 %, here it put 65 and said it’s urgent. So, okay, it already went down a bit in the level of urgency, right? Let me see here. Hi, how are you? I’d like to know the cost of the annual checkup and if she needs to fast for blood tests. Look , doesn’t need to be today, it’s a routine visit and it’s fine. So this one is correct and now it took very

17:07 close to what was promised there, right ? Well, I thought it was good, although it is health here, right folks, this issue, right? But I thought it was quite careful, right? Iara from Decisions API. Well, I’m going to test the photo feature too so we can see how it works. I’ll send the prompt, look. Now the owner can send a photo, create a separate screen that simulates the owner’s side, right, so we can test it as if it were a clinic contact form, fields for owner name, pet name, species, so everything they have to send and the photo, okay? Put the

17:38 access to this screen on the clinic dashboard and on the owner’s screen a path back to the dashboard. Upon sending, the message enters the dashboard inbox with the photo thumbnail and is analyzed right away. The owner’s screen shows a confirmation that the message has arrived. When the message has a photo, send the photo along with the text in the same call to the Decision API. Right? Because you can do it the way the documentation instructs to send messages. Resize the photo before sending so you don’t waste tokens unnecessarily. Add a fourth question. So we’re going to add this to the photo functionality, right? Only for messages with photos. If the photo shows any injury, bleeding, or visible

18:09 swelling, it’s also a yes or no question. With a percentage chance. Clinical rule: if that chance exceeds 70, the message severity level increases and the dashboard shows it rose because of the photo. Display the photo on the dashboard next to the answers with the tokens and the cost of that analysis, okay? Let’s grab a photo of a dog here. And guys, before we continue, if you’re thinking about getting a VPS, I wanted to recommend Hosker to you. That’s where we have all our applications hosted, okay? They have a special promotion going on here this month of October. And you can also

18:40 use the “hora de codar” coupons which will give you an extra discount on top of this promotion, okay? And many people have questions about choosing a plan. I recommend the KVM2 because it’s an excellent machine, you’ll be able to host several projects at the same time, okay? And with that, you can test several systems, including including the API, the dev, whatever you need, alright? Just come here and choose the plan. Then on this screen you’ll choose the duration. The longer the time, the cheaper it gets, right? They give more discounts, correct? Server closest to

19:10 you. These milliseconds here matter, right? But nowadays you can add a CDN if you want and practically zero out that value. Ubuntu is the distribution I choose for my VPS because it’s good for both beginners and more advanced users, okay? Ubuntu, forever, never had a problem. I confess Ubuntu is applied, so go for it, you’ll have your first VPS, alright? And of course, best cost-benefit, an excellent machine for any type of project. KVM2, okay? Let’s move on. Hey guys, so it ran here, okay ? Then we’ll have the tutor screen to

19:42 run the test with the image, okay? I’m going to see how it turned out here, look. It actually grabbed the image, man. Wow, now I’m stunned by the cloud. It’s all good. Let’s continue here. So, I’m going to access the tutor here. Then, I’ll enter my name. The pet’s name is a dog who was very happy with his last appointment. Thank you very much, guys. Some feedback, right? I’m going to send a photo of the dog. Look, here is the photo. I’ll send it. And

20:14 the message was received, and, you know , it arrived with the photo. It’s talking here, okay? So, I’ll check the dashboard here. There’s Zeus. Visible injury, bleeding, or swelling, no. So it is 100%confident in that, okay? Is it here to be seen today? No. But the message is different, okay? There is no feedback here, it’s all good. So that’s it, it worked. Image analysis is also a great differentiator of the Decisions API. The AI will probably fit this in, right? We take a normal request here, look, the time is practically the same.

20:46 Token usage is a little higher, 0.00301 value. Practically the same thing, right? A little more, a little more, but man, it’s negligible, right? There’s a lot. Oh, it’s even down here. I thought it wasn’t here. So, look at that. Oh, it even says it, right? A message without a photo, on average, spends this here, right? And with a photo, it spends that much there, okay? So, it uses a few more tokens. But, man , making decisions with images is very interesting, right? It opens up many paths too. So, a point for the Decisions API, right? Well, guys,

21:17 that’s what I wanted to present to you about the Decisions API, right? Just so we can talk about this release and you can also understand how to integrate it . This link here says it all, right? It’s the documentation, the reliable source of information. I suggest you read it here very carefully. Use this markdown file here, which is basically just adding .md to the URL, and it becomes documentation for you, okay? And you will have the same experience I had here, and I want to know what you thought of it, okay? Don’t forget, hosting is at Host, guys. If you want

21:48 to hire a VPS, they are on sale this month, using our coupon, right? The link will be here in the description. You will get a special discount to use and add to the discount they are already offering in this October promotion. And my recommendation is always the KVM2, because it is a machine with excellent configurations, you can put projects, several projects running at the same time, okay? With a very good cost-benefit ratio. You can compare how many hosting services they beat when you compare value and features, right? There’s no point in

22:18 having a cheaper VPS if it can’t handle your application, especially since we integrate with so many things; you need good CPU and RAM, or your application and system will crash, okay ? So go with CVM2 and you’ll be safe. And guys, another thing, right? They always invest in infrastructure here; they even have servers in Brazil today, so you can sign up right now for very low latency, which will make people access your system faster, right? Other hosting services are usually outside of Brazil. You can install several apps

22:49 right from the start, you know? You don’t need to waste time with anything. And there’s also the agent, which used to be called Codi; for me, it’s still called Codi to this day. With it, you can answer any questions you have about your hosting. It can help you host everything. It understands hosting . It’s an AI and it can manage your server for you, okay? You don’t even need to type commands manually. Choose the plan, make sure the coupon is applied to get that little discount, okay? Right here at checkout. The longer the term, the more discount you get. And here you choose a server in Brazil or the United States, right? To get lower latency. And also, Ubuntu is

23:21 the distribution I recommend, as it’s the best Linux for beginners, right? It’s very easy to work with and is a distribution that many others are based on. So, this setup is quite interesting , okay? Our Jeev course is also for those who want to learn about these decision AIs, understand how they work in depth, and see various use cases and projects; it’s all here, okay? The link will be in the pinned comment for you to check out my newest Jeev training. Well guys, I’m signing off here. I hope you enjoyed the integration of the decision API. It’s

23:53 quite simple to do, no mystery there. The time they promised is a little misleading, right? It responds a bit slower, but it’s barely noticeable, as you saw there, okay? It costs a bit more than Jeev and it accepts images, right? Which is a very interesting differentiator, but I believe Jeev will accept them soon too. Then the fight will be over price, if the response quality of both is the same, right? Well, if you guys want a video on JV comparing it with Decis DPI, leave it in the comments below and I’ll make it for you too, and we can see which one performs better based on the evaluation

24:23 criteria, which is what matters. Time for time, it’s practically the same for the person waiting, and the price won’t make much of a difference either, alright guys? I’ll wrap things up here, see you in the next video or over in Jeff’s course as well, sound good? Thanks, see you later. M.