With the release of o1, OpenAI’s first model in a new line designed for deeper reasoning, let’s use this thread to share tips, tricks, and best practices. Whether you’ve found ways to optimize performance, improve accuracy, or enhance results for specific tasks, feel free to contribute! This can be a great resource for developers looking to get the most out of o1 in real-world scenarios.
5 Likes
WE DONT HAVE IT YET IN OUR GPT PLUS 
I got access a few minutes ago.
Hang in there!
4 Likes
Does this new model have access to memory like gpt-4o does?
This is from the writeup that cognition did on using o1:
"Prompting o1 is noticeably different from prompting most other models. In particular:
-
Chain-of-thought and approaches that ask the model to “think out loud” are very common in previous generations of models. On the contrary, we find that asking o1 to only give the final answer often performs better, since it will think before answering regardless.
-
o1 requires denser context and is more sensitive to clutter and unnecessary tokens. Traditional prompting approaches often involve redundancy in giving instructions, which we found negatively impacted performance with o1.
I think everything that was used in the past works just as well now too. It may take a bit longer to get back a response, but the responses as much better now!
Awesome topic. One time I was doing many workaround till see the OpenAI recipes. I hope this topic grow.
This is of course bot nonsense.
1 Like
What about Python coding performance compared to gpt-4o-2024-05-13 ?
If someone has information this could help other, thank you !
1 Like
(this pricing task is because the above originally also asked about pricing)
Python coding? Let’s try a Python coding request on ChatGPT’s GPT-4o as a baseline:
Complete non-compliance with the task instructions. There was no Python script. The input message may have been truncated without error by ChatGPT, or the model didn’t comprehend.
o1-preview follows the task instruction:
However “o1-preview” included only o1 and gpt-4o models in its python listing, despite a massive comprehensive input of all models and no failure on it:
To answer a question about pricing, we use the resulting script to see if it works:
Select a model from the following list:
- gpt-4o-2024-08-06
- gpt-4o-2024-05-13
- gpt-4o
- chatgpt-4o-latest
- gpt-4o-mini
- gpt-4o-mini-2024-07-18
- o1-preview
- o1-preview-2024-09-12
- o1-mini
- o1-mini-2024-09-12
Enter the number corresponding to your model choice: 6
Enter the number of input text tokens: 34749
Enter the number of output text tokens: 1846
Enter the number of images: 0— Pricing Summary —
Model selected: gpt-4o-mini-2024-07-18
Total input tokens (text + images): 34749
Total output tokens: 1846
Input cost: $0.005212
Output cost: $0.001108
Total API call cost: $0.006320— Pricing Summary —
Model selected: o1-mini-2024-09-12
Total input tokens (text + images): 34749
Total output tokens: 1846
Input cost: $0.104247
Output cost: $0.022152
Total API call cost: $0.126399
That’s not the real answer for o1. The much higher actual cost can only be guesstimated.
What this doesn’t consider is o1 using many internal context-building AI calls to reach an answer, all at your expense, so the cost could be over a magnitude greater. (Knowing that is why I let ChatGPT do it).
ChatGPT progress report of a follow-up task to fix, lots of reasoning thought about a price list longer than just chat models
Organizing model types
The task involves revising a Python script to include all models, sorted by type, and ask for the model type and name in succession. This approach ensures the script efficiently handles different model categories.
Pulling together details
Gathering the models from the HTML document, classifying them by type, and checking their vision support to craft a tailored script.
Categorizing model types
I’m organizing models into sections: GPT-4o, GPT-4o mini, OpenAI o1-preview, OpenAI o1-mini, Embedding models, Fine-tuning models, Assistants API, Image models, Audio models, and other Models.
Organizing model types
I’m working through organizing models by type and vision support, updating the script to prioritize these categories. Extracting details from HTML to streamline the process.
Mapping out models
I’m thinking through different GPT-4o models and their pricing. The latest model, GPT-4o-2024-08-06, supports vision at $5.00 per million tokens for input and $15.00 for output. Interestingly, GPT-4o-2024-05-13 and gpt-4o charge the same for input and output tokens.
Weighing model benefits
GPT-4o mini is a cost-effective, vision-capable model. It compares favorably to GPT-3.5 Turbo, making it a smart choice for budget-conscious users.
Introducing o1-mini
I’m learning about OpenAI’s o1-mini, a cost-effective model for coding, math, and science. It’s priced at $3.00/1M for inputs and $12.00/1M for outputs. No vision support is mentioned.
Analyzing Fine-tuning models
I’m digging into various fine-tuning options for text and input/output prices. Notable models include GPT-4o and GPT-4o mini, which offer free fine-tuning up to a daily token limit through September 23, 2024.
Evaluating relevance
OK, let’s see. The Assistants API, image, and audio models don’t align with our script’s needs. Progress is being made by focusing on pertinent elements.
Listing model options
I’m preparing a detailed list of supported models by grouping them by type and noting vision support. This includes various GPT-4, GPT-3.5, and Davinci models with specific input and output prices.
Mapping model features
I’m exploring models by type, pricing, and vision support. GPT-4o and GPT-4o mini excel in vision, while others likely lack this capability. Confirming these features helps in understanding their capabilities.
Mapping vision support
OK, let me see. The FAQ lists GPT models with vision support, so I’m piecing together a categorized, vision-support-focused summary.
Model selection and cost options
I’m starting to map out the different models offered, each with details on pricing, tokens, and vision support. It’s interesting to see how these options provide a spectrum of design possibilities.
Weighing vision models
I’m considering if the o1-mini models support vision, but since they don’t specify, it’s safe to assume they do not. Including embedding models is also necessary.
Evaluating model options
I’m breaking down the models offered by OpenAI, detailing their pricing and whether they support vision tasks. Notable models include text-embedding-3-small and gpt-4o-2024-08-06.
Reviewing model specifics
I examined various models, focusing on their specific capabilities and pricing. Training costs were deemed irrelevant for this script, which calculates only usage costs.
Mapping out model support
I’m gathering details on vision-enabled models. The supported ones include GPT-4-2024-04-09, GPT-4-Turbo, and others, organized by model type, pricing, and vision support.
Breaking down models
I’ve mapped out the models into a nested dictionary by category, detailing input/output costs and vision support. This structure aids in understanding costs clearly.
Crafting the script
Starting with model types, then narrowing down to specific models. Including vision support and embeddings. Ensuring clarity and thoroughness throughout.
Breaking down model support
I’m listing model types and their attributes like pricing and vision support. This detailed structure aims to guide implementation accurately.
3 Likes
I have noticed that asking the model to be thorough typically results in longer thinking time and usually better results, Outputs will be much longer though.
2 Likes