On January 15, 2026 a guide by Hillary Bush and Prashant Mital lays out practical steps for startups to adopt the GPT-5 model series. The authors map a four-step process—migrate, optimize, steer, and troubleshoot—and report patterns observed while working with dozens of early- and growth-stage companies.
Migrate to the Responses API
The guide recommends moving workflows to the Responses API to take advantage of GPT-5’s reasoning features. According to the authors, only the Responses API lets the model persist “reasoning items” across turns and tool calls, giving each request access to its internal context. That persistent context is said to improve performance, increase cache-hit rates, and help lower infrastructure costs. The guide also highlights compatibility with zero-data retention (ZDR) as a deployment option.
Optimize prompting
The authors advise beginning with existing evals to establish a baseline and then iterating. They recommend streaming reasoning summaries for specific failure cases to see where the model’s internal steps diverge from expectations. The guide notes GPT-5 is effective at metaprompting—using the model to refine its own prompts—and suggests keeping instructions shorter and clearer when possible. When prompts are stable, teams should capture them in templates and document examples of good and bad outputs for consistent reuse.
Steer GPT-5 and troubleshoot common issues
New controls described in the guide include reasoning_effort (options: minimal, low, medium, high; default: medium) to adjust how much the model “thinks”, and verbosity (low, medium, high) to influence output length. The authors stress systematic experimentation against evals to find the right configuration. Common problems identified by the authors include overthinking, underthinking, over-deference, excessive verbosity, latency, tool overuse, and malformed tool calls; many are addressed through prompt tuning, metaprompting, and targeted diagnostics.
The guide cites a reported GPT-5 performance figure of 74.9% on SWE-bench Verified and recommends teams migrate to the Responses API to access the model’s reasoning and control features.
Original source: OpenAI News