o3 pro stands out for its ability to handle complex requests. Photo: OpenAI. |
OpenAI has launched o3 pro in its Pro package for $200/month and Team via API. An upgraded version of the o3 that was introduced a few months ago, o3 pro is considered by the company to be the most powerful one yet.
All versions with the “pro” tag will be associated with the ability to answer more difficult and longer questions. Unlike regular versions of AI, inference models solve problems step by step, allowing them to work more reliably and consistently in areas such as physics, mathematics, and programming.
“We recommend the o3-pro for difficult questions where reliability is more important than speed, and waiting a few minutes is worth the trade-off,” the company said. In shared test reviews, the o3-pro outperformed the o3 and o1-pro versions.
Ben Hylak, a former Apple employee and co-founder of AI startup Raindrop, said the new model is much smarter. He gathered a history of all previous meetings at his company and then asked o3-pro to create a plan.
The results were impressive, as specific and analytical as he had hoped a large language model (LLM) would be. The plan included target metrics, timelines, priorities, and strict instructions on what to eliminate completely. “It was so specific and grounded that I had to rethink the future of my company,” he writes.
The results from the o3 pro (left) are more specific and grounded. Photo: Ben Hylak/X. |
O3-pro costs $20 per 1 million input tokens and $80 per 1 million output tokens when used via API. That's the AI's ability to remember and process data. One million input tokens is equivalent to about 750,000 words, which is longer than the book War and Peace , The Verge compared.
OpenAI writes that experts consistently gave the o3 pro a higher rating than the o3 in every category tested. Evaluators also gave the o3 pro a higher rating for consistency across a number of criteria, including clarity, ease of following instructions, and accuracy, especially in key areas like science,education , programming, business, and writing support.
On AIME 2024, a test that assesses a model's mathematical abilities, the o3 pro scored higher than the Gemini 2.5 Pro, Google's flagship AI. In addition, the model also surpassed Anthropic's Claude 4 Opus on GPQA Diamond, a test of doctoral-level scientific knowledge.
o3 pro also comes with built-in tools that allow it to search the web, analyze files, use Python for calculations and programming, and personalize responses by leveraging memory. Commenting on this aspect, Ben Hylak said that the tool clearly demonstrates its ability to recognize its surroundings, when to ask about the outside world (instead of pretending to know), and choose the right tool for each task.
However, the model’s biggest drawback is its response time, which is even slower than the o1 pro. YouTuber Bijan Bowen agrees. “Although the model’s answer is quite clear, in just a few sentences of description, the response time is quite long,” he said. Especially in cases where there is not enough external data, the model tends to think randomly, Ben Hylak added.
The O3-pro also has some other limitations, such as the inability to generate images, as well as support for the Canvas feature. ChatGPT's temporary chat feature with this model is currently disabled while OpenAI fixes a "technical issue."
However, Mr. Hylak said that this is not a model for users to chat like Claude 3.5 Sonnet or ChatGPT 4o. Advice from Nate B. Jones, Head of Product at Rockerbox, said that o3 pro should be used for difficult tasks that require 15-20 minutes of thinking.
Source: https://znews.vn/mo-hinh-suy-luan-moi-cua-chatgpt-post1560084.html
Comment (0)