OpenAI has just released its new reasoning model GPT-5 and expectations are high due to its capacity for complex tasks and its potential in software development. In this article I translate and summarize a real usage test by developers that I conducted using OpenAI's Responses API in a personal project called mycaminoguide and in a standalone application that analyzes customer data extracted with Airbyte Embedded MCP.
Important notice: these tests are neither scientific nor exhaustive. I was looking to measure real and average usage by a developer in specific use cases.
Test 1 end to end: the application is written in Python with a Streamlit front end and uses Airbyte Embedded MCP to fetch Stripe invoices on behalf of a customer. The flow consists of obtaining a Bearer token, calling the MCP to bring the data, and then passing that data to the Responses API so the model can analyze it. The main code is around 114 lines. When running three times with GPT 4.1 the times were 29.85 seconds, 15.38 seconds, and 13.56 seconds. When changing only the model to GPT 5 the times were 77.72 seconds, 57.72 seconds, and 83.40 seconds. These results were surprising because GPT 5 was much slower in this scenario of chaining MCP and analysis with the model.
Test 2 removing variance: to ensure the delay was not coming from MCP or Stripe calls, I separated the MCP call and then injected the data directly into the prompt. I kept timers for the proxy call and for the OpenAI call. With GPT 4.1 the proxy calls took 6.71, 6.74, and 6.93 seconds and the OpenAI call took 9.87, 10.31, and 9.62 seconds. With GPT 5 the proxy calls were 7.85, 7.01, and 7.11 seconds while the OpenAI call took 80.01, 46.05, and 48.06 seconds. In other words, the latency attributable to the model using the Responses API with GPT 5 was dramatically higher and the proxy call showed little variation, suggesting the problem was during the interaction with the API and the model.
Test 3 the Responses API introduces variance: to rule out that it was an API flow issue, I used the same data string and tested the same prompt directly from ChatGPT with GPT 4o and GPT 5. The prompt requested an invoice analysis and that the model report exactly how long it took in seconds with two decimal places. With ChatGPT the times reported by GPT 4o were 4.92, 5.32, and 3.47 seconds. With GPT 5 the times were 38.82, 36.78, and 34.62 seconds. Observation: ChatGPT seems to measure time including the generation of the response on screen, so the times printed by the model itself may include the visual typing phase, but the magnitude difference between 4o and 5 remains notable.
Provisional conclusion: in my developer tests and in real integration scenarios with customer data, GPT 5 behaved significantly slower than GPT 4.1 and GPT 4o through the Responses API. I have not compared output quality in detail; it is possible that GPT 5 offers deeper reasoning that requires more time. Another hypothesis is that GPT 5 is suffering higher usage load at the time of testing. Open questions remain such as comparing GPT 5 with open source versions or GPT 5 mini to filter server loads and see if the behavior persists.
At Q2BSTUDIO as a software development and custom applications company we remain very attentive to these advances. We are specialists in custom software, custom applications, and applied artificial intelligence for businesses. We offer comprehensive services including cybersecurity, aws and azure cloud services, business intelligence services, and power bi solutions. We work on integrating AI agents and AI solutions for companies that bring real value to decision-making and automation processes.
If you are a company and are interested in improving your processes with artificial intelligence, we can help you evaluate models based on performance, cost, and quality criteria. In projects where latency is critical, we can propose alternatives such as prompt optimization, batching, use of lightweight models, or private deployments to reduce response times. Our team masters the integration of AI agents, power bi for visualization, and secure data pipelines in the cloud with a focus on cybersecurity.
Practical recommendations after these tests: 1) always measure latency in real production scenarios before choosing a model for the application, 2) consider cost-benefit if the improvement in reasoning justifies the higher latency, 3) test alternative models and architectures that decouple data fetching and model evaluation to control variance, and 4) monitor usage and latency to detect load spikes that affect the service.
At Q2BSTUDIO we are available to advise on custom development projects, artificial intelligence integration, secure deployments in aws and azure cloud services, business intelligence projects, and cybersecurity solutions. If you want to build AI agents, improve processes with AI for companies, or create interactive reports with power bi, we can collaborate to design a scalable solution adapted to your business.
For now I will continue testing and refining. In the short term I will continue developing with GPT 4 to maintain acceptable response times and reserving tests with GPT 5 for cases where greater reasoning depth can justify the latency. If you are interested in us conducting similar tests on your infrastructure or evaluating models for your use case, contact Q2BSTUDIO and let's talk about custom applications, custom software, and applied artificial intelligence.
