Application Cost Overview
A Pre-Deployment Evaluation of LLM and API Operational Costs
The four workflows of the application are now fully operational, and the official demo video will be released shortly. This note outlines the costs associated with each of the four workflows when powered by GPT-4o-mini. The token usage and corresponding cost estimates are based on the OpenAI usage report.below:
Tokens counts and costs estimations are based on the OpenAI usage report.
1. Articles search workflow
This workflow is powered by two free APIs: arXiv and CrossRef. Built-in throttling mechanisms ensure API rate limits are not exceeded. Query interpretation is handled by the LLM.
Token usage: ~8,000 tokens
Cost per request: ~$0.03
Test query: "pair-trading"
2. Summarisation (Insights extraction) workflow
The source text is pre-processed using NLP techniques to minimize token usage and reduce input noise before being summarized by the LLM.
Test article:
Faber, Meb. “A Quantitative Approach to Tactical Asset Allocation” (2013). The Journal of Wealth Management, Spring 2007. SSRN LinkToken usage: ~8,500 tokens for a 17-page article
Cost per summary: ~$0.03
3. Coding workflow
Code generation is currently for a code of around 60 lines.
Token usage: ~4,500 tokens
Cost per generation: ~$0.02
4. Chat with Fundamentals
This workflow leverages three APIs from EODHD:
⚠️ Note: GPT-4o-mini is currently not compatible with the data model used in this workflow. As a result, GPT-4o is used instead for the time being. While more expensive, it provides a higher quality of analysis.
Test query: "Is TSLA a good buy?"
Token usage: ~10,000 tokens
Cost per query (GPT-4o): ~$0.05
API calls to EODHD: ~420 per request
To provide perspective, EODHD’s paid plans allow up to 100,000 API calls per day, while the free tier allows only 20 calls per day, making it insufficient to run this workflow in its current form.
5. Side Costs
Platform is deployed as local-first model and deployment-related expenses are not expected at this stage. Looking ahead, the ultimate objective is to evolve the system toward a locally hosted version powered by high-performance computing and local LLMs.
Best regards,
SML