OpenAI officially launched its new AI model, 'OpenAI o1' which was previously known as 'Project Strawberry'. Marking a significant leap in AI development. The o1 model is designed to handle complex tasks with improved reasoning capabilities, excelling in fields like science, coding, and math.
OpenAI o1 focuses on 'thinking before responding,' improving problem-solving by refining its reasoning process and recognising mistakes. It demonstrates enhanced performance in multistep challenges, especially in generating and debugging code. In benchmark tasks like advanced math, physics, chemistry, and biology, OpenAI o1 performs at the level of PhD students.
Open AI said in their offiicial blog "OpenAI o1 series models are new large language models trained with reinforcement learning to perform complex reasoning. o1 models think before they answer, and can produce a long internal chain of thought before responding to the user. o1 models excel in scientific reasoning, ranking in the 89th percentile on competitive programming questions (Codeforces), placing among the top 500 students in the US in a qualifier for the USA Math Olympiad (AIME), and exceeding human PhD-level accuracy on a benchmark of physics, biology, and chemistry problems (GPQA)."
How reasoning works
The o1 models utilise 'reasoning tokens' to enhance their problem-solving abilities. These tokens allow the models to 'think' by analysing the prompt and evaluating various approaches before generating a response. After this process, the reasoning tokens are discarded, and the model produces a final answer using visible completion tokens.
Here’s an example of a multi-step interaction: input and output tokens are retained throughout the conversation, while the reasoning tokens are removed from the context.
/socialsamosa/media/media_files/7V3KkRjWin5I3FbYeP6L.png)
The model lacks several features present in the GPT-4o model, such as web browsing, and the ability to process images or files. For now, it only handles text-based prompts.OpenAI CEO Sam Altman acknowledged that the o1 model is still limited, slower, and imperfect, as it prioritises thoughtful problem-solving over only speed.For common tasks, GPT-4o may remain more capable in the near future.
OpenAI also released a smaller, more affordable version called 'OpenAI o1-mini,' which is faster but still retains the model's reasoning capabilities.This version is tailored for developers who require efficient coding solutions. OpenAI o1 is available to ChatGPT Plus and Team users, who can manually select between o1-preview and o1-mini in the model picker. Weekly rate limits are set at 30 messages for o1-preview and 50 for o1-mini, with plans to expand access.
OpenAI o1-mini is expected to be made available to free users soon, while o1-preview comes at a higher cost. In the API, o1-preview costs $15 per 1 million input tokens and $60 per 1 million output tokens.Comparatively, GPT-4o is priced at $5 per 1 million input tokens and $2.50 for output tokens. In India, accessing OpenAI o1 costs around Rs 1,260 per million input tokens, significantly higher than GPT-4o's Rs 420 price.The ChatGPT Plus subscription, which includes o1 access, costs $20 per month (Rs 1,650 in India).
/socialsamosa/media/agency_attachments/PrjL49L3c0mVA7YcMDHB.png)
/socialsamosa/media/media_files/2024/10/21/leaderboard-new-alt.png)
Follow Us/socialsamosa/media/media_files/rIoiANbbTA88hFT1MNyj.jpeg)