New tools for building agents: Responses API, web search, file search, computer use, and Agents SDK

This is the case for me as well. Did you move your account to tier 3 after the model’s release?

1 Like

Yes, I did! I was wondering if that had any impact

Yes, please we need this in Typescript.

2 Likes

I also get the same problem and my organization is on Usage tier 3.

Curiously, yesterday, computer-use-preview appeared in my /models endpoint response (which prompted me to try it out), but a few hours later when actually I try to use it, I get the same “The model computer-use-preview-2025-03-11 does not exist or you do not have access to it.” error, and when I look at /models endpoint again it’s gone :face_with_diagonal_mouth:

1 Like

Any updates on computer use?
Someone able to test it?
If not, does someone know when its available?
I topped up credits to get to Tier 3 and am eager to try it.

1 Like

The Assistants API, which will be deprecated in a year, had functionality to create a run that will happen asynchronously and then query for the run result using the run ID.

Is this possible with the Responses API? As far as I could tell, it was not immediately obvious how to do that if possible.

Many of us are running in serverless environments (Vercel/lambda) and the cost associated with waiting on a response would be orders of magnitude lower if we didn’t have to do that, and that’s putting aside the point that for hobbyists some providers don’t even allow for a long enough time out.

Hello,

Do you have an ETA for audio output support in the Responses API?

Thank you in advance for your reply!

1 Like

I seem to be having the same problem as others are having and still do not have access to this despite having upgraded to tier 3 earlier today, would be good to get a timeline on when this will be properly rolled out

I would just go with the GPT 4.5 preview API and add the computer use, internet use, and other agentic functions directly into your model. You’ll get results! Start small, then once you get results you can slowly add more functions.

computer-use-preview is its own model https://platform.openai.com/docs/models/computer-use-preview

This means there are undisclosed criteria holding it back from others.

Can you still use web search with 4.5? It was working for me before yesterday and now it says web search is not supported.

Indeed, I overread the “select” part, this should be bold, underlined, and red lol

spending money doesn’t give you access to the CUM


So this means we can get access to it in 3-6 months, maybe :laughing:

Sick! I have made Agent SDK in Go Lang.
It´s @pontus-devoteam/agent-sdk-go at GitHub.

Would be awesome to see it under your community libraries

2 Likes

These are some great updates in my opinion.

I dunno if anyone noticed, but the way the File Search and Vector Stores are now working—e.g. being able to be called directly by API—makes the OpenAI platform a good choice for any 3rd party Vector Database. Good move.

@edwinarbus, I read in the Agents SDK section on Models that GPT 4.5 is “best for agentic execution.” Everything I’ve heard about 4.5 is that it’s better at creativity and writing, so hearing it’s best for being your new agent model is news to me. Do you (or anyone) know why it’s best (or better) for Agentic actions?

This seems like a truly understated super power given the mild hype surrounding the model’s release.

I agree this doesn’t make much sense. GPT-4.5-Preview in Agents?

  • Take a model that is $75.00 / 1M tokens input, thus $7.50 for 100k of 125k context.
  • Let it loose to call or be called over and over.

It only makes sense in that this is already a model to check back in with later, because of its lower generation rate.

The biggest shortcoming that may make it not suitable is that an output over 2000 kilotokens is almost impossible to get generated, and the AI will not be coached into making more - finding a place for you to do the rest of the work regardless, and offering even less when you point out the well-justified “expanded version” was 1700/16384 allowed.

So I would strike that from documentation. GPT-4.5 could be called for a very specific purpose by an orchestrator.

(BTW: this is not to say this model can’t be amazing. It is able to, without producing planning, completely refactor code into a new structure with all new methods, where o3-mini will just add more pasta sauce)

1 Like

Hidden Costs of Web Search

I have just set up the Responses API using web_search_preview tool, I have selected model gpt-4o, but in my usage dashboard, I seem to have been charged for gpt-4.5-preview-2025-02-27, output, gpt-4.5-preview-2025-02-27, input and web search tool calls | gpt-4o, med

This isn’t clear from the docs that 4.5 would be used, this has to be an error right?

Please give us typescript package

Been waiting for this. Thanks so much!!

Looking at the Web search API, when city, region and country fields are used, are web searchs conducted in the specific country - for example “FR” (France)? If so, would the language of the response be the language of the country field (ISO country code)?

Whatever the case, it should be discussed in the Web Search API section.

I set the search engine language fields to Paris, France. The region selected does have an impact on the results, from prior experience.

An analysis by the AI employing the tool (and made one purpose, to help diagnose internal tools) then shows that results are in the query language, not the region.

Upon conducting a web search for “shopping egg prices,” the following results were obtained, all presented in English:

  1. US turns to Brazil for eggs and considers other sources during bird flu outbreak

    • Language: English
  2. New Yorkers scramble for free eggs

    • Language: English
  3. Egging houses is just too expensive these days

    • Language: English
  4. Rising egg costs scramble Philly’s breakfast scene

    • Language: English
  5. Can suspending a cage-free egg law solve the soaring price problem? Nevada takes a crack at it

    • Language: English
  6. ‘My job is basically just eggs now’

    • Language: English
  7. Egg prices could jump 41% this year, USDA says, as Trump’s bird flu plan is unveiled

    • Language: English
  8. :cooking: Egg crackdown

    • Language: English
  9. Explore Fresh Egg Prices, Forecasts & Historical Data

    • Language: English
  10. Higher egg prices drove up inflation, and Walmart grocery revenue

    • Language: English
  11. Explore Egg Prices, Forecasts & Import/Export Data

    • Language: English
  12. How Much Do Eggs Cost? Factors Affecting Egg Prices, Types, And Savings – CookYourTech

    • Language: English

Next, translating the query “shopping egg prices” into French yields “prix des œufs en magasin.” Conducting a web search with this French query resulted in the following outcomes, all presented in French:

  1. Prix des œufs en France : tendances et prévisions

    • Language: French
  2. Les prix des œufs atteignent des sommets en raison de la grippe aviaire

    • Language: French
  3. Comment la grippe aviaire influence le marché des œufs en France

    • Language: French
  4. Comparatif des prix des œufs dans les supermarchés français

    • Language: French
  5. Les consommateurs français face à la hausse des prix des œufs

    • Language: French
  6. Impact de la grippe aviaire sur les producteurs d’œufs en France

    • Language: French
  7. Les œufs bio sont-ils plus chers en période de crise sanitaire ?

    • Language: French
  8. Stratégies pour acheter des œufs à moindre coût en France

    • Language: French
  9. Analyse du marché des œufs en France en 2025

    • Language: French
  10. Les alternatives aux œufs face à la hausse des prix

    • Language: French

In summary, the initial search with the English query returned results exclusively in English, while the subsequent search with the French-translated query yielded results entirely in French.

That is just for the search engine. If it were to click through, is the AI going to be reading English again? We have to use the tools again since the context paid for is not supplied.

Based on the search results for “prix des œufs” (price of eggs), here are the answers to your questions regarding the quality of the results:

  1. Are the site URLs, given in the results, ones that you can then make a follow-up click on with a method offered to you in the internal tool methods you can send to obtain just one site’s web page data?

    Yes, the URLs provided in the search results are accessible and can be used to retrieve individual web page data. For instance, the article from Le Monde titled “L’inflation, si elle se matérialise immédiatement par la hausse des prix, ses cicatrices sont longtemps visibles et douloureuses” is available at https://www.lemonde.fr/economie/article/2024/11/08/l-inflation-si-elle-se-materialise-immediatement-par-la-hausse-des-prix-ses-cicatrices-sont-longtemps-visibles-et-douloureuses_6383253_3234.html. This URL can be used to access the specific web page data.

  2. Are the URLs offered in the results specifically for sites that would be in the French language?

    Yes, the URLs correspond to websites that are in the French language. For example, the article from Le Monde is in French, as indicated by the content and the domain “.fr”, which is the country code top-level domain for France. Similarly, other sources like https://rennes.maville.com/actu/actudet_-le-prix-des-oeufs-pourrait-bientot-augmenter-a-cause-de-la-hausse-des-couts-de-production-_54135-5025940_actu.Htm are also in French.

Therefore, the search results are appropriate for follow-up actions using internal tools to obtain specific web page data, and they are in the French language, aligning with the original query.

So, since you also have a helpful AI, you can continue to answer such a question - whether the sites themselves are localized and going to be by the query language or the parameter selected.

1 Like