I’ve been using the “Deep Research” feature in ChatGPT and find it incredibly powerful for synthesizing complex information and extracting insights from long or technical documents. I’m now looking to build a similar capability into an automated workflow using the OpenAI API.
Does the OpenAI API provide access to the same underlying model or capabilities used in the “Deep Research” feature in ChatGPT? If not, what are the limitations or reasons? Also, are there best practices or architectural suggestions to replicate Deep Research-like functionality using the API (e.g., document chunking, RAG, long-context handling)?
It is an application-specific fine-tuned version of o3.
Additionally, Deep Research is able to make dozens of web search crawls of a more exploratory nature than the internal web search tool offers.
You could prompt an AI to “dive deep” into your own paid API search service as tool, and web page scraper. However, with the output hesitancy limitations of API models, being resistant to writing complete Wikipedia articles for you (while my most recent “Deep Research” was 4000 tokens output for even a recipe reformulation), to be able to offer such a long-token report would also be a hurdle.
One could guess at the reasoning: high cost per use, but which as an enticement to upgrade the subscription level of ChatGPT, can be amortized over the number of actual uses someone might make within a $2400/yr ChatGPT Pro subscription for more than a handful per month.
There is a fallback version: use up the big daddy for the month (10 per Plus), you can still get the same kind of service with a lesser model. That is the o4-mini.
Once you reach your limit for the full version, your queries will automatically switch to the lightweight version
Plus, Team, Enterprise, Edu – 10 tasks/month, plus an additional 15 tasks/month using the lightweight version
Be careful—sometimes he mixes the information just to please you and is feeding you with BS.
Looks nice at first glance, but when you go deeper, follow his trail, and check what he did, you will be disappointed.
It’s not rocket science to do the same orchestration (if you check on the internet, there are even examples with Ollama).
