What OpenAI's recent models o1-preview and o1-mini mean for developers

OpenAI surprised the world yesterday afternoon by announcing neither “Strawberry,” as rumored, nor GPT-5, but a brand new family of “logical” large language models (LLMs) called o1, designed to supply high performance and accuracy on science, technology, engineering, and arithmetic (STEM) tasks.

OpenAIs two recent models are o1-preview and the less advanced o1-mini, which is now available for ChatGPT Plus users and developers who Paid application programming interface (API) from OpenAIThis allows developers to check them because the backend of existing third-party apps and services or to construct recent apps and services on top of them.

The recent o1 models use a type of “argumentation,” OpenAI says, and so they “try different strategies, detect errors and undergo the entire thought process,” said Michelle Pokrass, OpenAI’s API tech lead, who explained a few of the pondering behind the models’ development in a video interview with VentureBeat.

“In our tests, these models perform quite similarly to PhD students on a few of the most demanding benchmarks,” Pokrass noted.

Especially for “reasoning-related problems,” the o1 models perform “significantly higher” than the GPT series, said Nikunj Handa, who works in product at OpenAI and in addition took the time to share his thoughts on the o1 model family with VentureBeat.

Here’s what third-party developers should know concerning the recent o1-preview and o1-mini models.

Limited to text – no image or file evaluation – and slower… for now

The o1-preview and o1-min models are currently limited to text inputs and outputs and thus are unlikely to exchange third-party use of GPT-4o at this stage, OpenAI's last and most advanced model that provides multimodal inputs and outputs including file attachment evaluation and image generation.

According to Pokrass and Handa, the O1 series models usually are not multimodal.

The o1 models also cannot yet hook up with the Internet and due to this fact don’t have any external knowledge after the training cutoff date (October 2023). However, users can after all provide their very own knowledge to the model in the shape of text inputs in order that it may possibly draw on it and analyze it.

They are also slower to reply with results; in some cases it takes greater than a minute – sometimes even several – to get a response.

However, some developers who’ve received early alpha access in recent weeks and months have reported improved performance in tasks similar to Coding And Preparation of legal documentsdue to this fact, using considered one of these systems could still be a great option for developers who wish to experiment and pay more for higher performance.

As OpenAI in its API documentation for its recent reasoning models o1-preview and o1-mini: “For applications that require image input, function calls, or consistently fast response times, the GPT-4o and GPT-4o mini models are still the best alternative. However, if you must develop applications that require deep reasoning and enable longer response times, the o1 models could possibly be a wonderful alternative.”

o1 costs way more than other OpenAI models, but o1-mini is a bargain

First of all, you will need to be an intensive user of OpenAI's APIs to qualify. The o1-preview and o1-mini models might be made available initially for Tier 5 users – that’s, those that spent $1,000 through the API and made payments to the corporate at the very least 30 (or more) days ago.

OpenAI warns that the brand new o1 models are previews and limited as much as 20 Requests per minute — or 20 calls per minute — in comparison with other OpenAI models which have higher limits or limited by tokens per minute/day.

The company also doesn’t currently accept “bundled” requests, as other lower-priced models do, which essentially bundles inputs into the API that don’t require a direct response, but slightly are analyzed and provided appropriate responses inside 24 hours (or less).

The primary o1 preview model, which Pokrass says offers much more “world knowledge” on topics outside of STEM, is by far OpenAI's most costly AI model currently offered, costing $15 per 1 million tokens input and $60 per 1 million tokens issued ($15/$60) versus $5/$15 for GPT-4o. So the brand new full o1 preview model is 200 to 300 percent costlier.

Still, the o1-mini model is a bargain at $3 per 1 million input tokens and $12 per 1 million output tokens, which is 80% cheaper.

“Of course, we’ll adjust prices in the approaching weeks and months to bring them to the best level,” said Pokrass.

Here is a breakdown of the costs of OpenAI’s various leading models via its API – data taken from this page.

When it involves context – or what number of tokens a specific LLM can process in an interaction, input and output – The o1 series has a limit of 128,000comparable to GPT-4o and the opposite top models from OpenAI.

The o1-preview model can produce a maximum of 32,768 tokens in a single output or response, while the o1-mini model can produce twice that number, namely 65,536.

What developers are using OpenAI o1-preview and o1-mini for to date…

It's been lower than 24 hours since OpenAI released o1-previews and o1-mini, but already some developers are pondering up uses for it and testing it to see what it does well and what it doesn't.

And as mentioned, OpenAI has been “spreading” it out to a select group of early alpha users and testers over the past few weeks and months.

Based on this work, listed below are a few of the most interesting uses of the o1-preview and o1-mini models to date:

Creating plans and white papers

Several users have reported that the model family o1 well-developed motion plans and even complete documents similar to Whitepaper with quotes based on easy prompts.

Planning, infrastructure and risk assessment

AI influencer and business consultant Allie K. Miller posted a thread on X with various impressive results from OpenAI's o1-preview model, including robotically (and far faster than a human) Optimizing human resource scheduling for a corporation that assesses Merger risks, Design of warehouses for efficiency, also Balancing a city’s power grid.

Quickly create apps and games

OpenAI o1-preview appears to be a direct warning shot to Anthropic's Claude family and particularly to the Artifacts feature, as additionally it is a strong and fast way for users to create their very own interactive apps and games, as Ammaar Reshi, Head of Design at AI speech and audio startup ElevenLabs, pointed to XNote that he used one other software tool, Cursor Composer, to run the model.

But as Anand Sukumaran, CTO of web notification startup Engagespot, posted on his X-accountGPT-4o still achieves much higher speeds when coding easy programs, similar to one which displays “Hello World!”.

Independently completing requests for proposals (RFPs)

Contractors, especially those providing products to government agencies, are all too conversant in the Request for Proposal (RFP) – an agency tender for contract offers in a standardised format that may be tedious and time-consuming to finish.

While specialized and AI-driven software has come onto the market to assist contractors complete these documents more efficiently, says Ethan Mollick, a professor on the Wharton School of Business on the University of Pennsylvania, a number one AI influencer and early adopter who had access to o1 as a part of the alpha testing phase, posted on X that o1 can populate RFPs itself – nevertheless, that is after all limited to text and doesn’t accept file uploads, so the user must copy and paste the text version of the RFP into o1's context window in ChatGPT or via one other app.

Engagement and growth hacking strategies

Ruben Hassid, founding father of EasyGen, a Chrome app for robotically generating LinkedIn posts, published a Demo video on X shows how o1-preview was capable of develop a comprehensive and well-thought-out plan to make use of Reddit to grow their business.

“I can't imagine how long the answers are. There's no way an LLM could cover that much strategy,” he wrote.

Where can I get access to OpenAI o1-preview and o1-mini?

Developers can after all access the brand new OpenAI o1 models via the company's public APIand thru Microsoft Azure OpenAI Service, Azure AI Studio, and GitHub models.

While that is clearly not for all (or possibly even most) developers, the debut of the O1 family is an exciting time for anyone trying to have room to experiment and create recent apps and services.

OpenAI can also be committed to advancing the capabilities of each the O1 family and the GPT series, so there isn’t any shortage of options for those trying to construct on the leading AI company's platforms.