The context · March 2026
OpenAI introduced GPT-5.4 mini and nano in March 2026, expanding the range of model sizes available for different tasks. [1]
Different jobs need different evidence
Classifying a short request, extracting a known field and resolving an ambiguous technical question are different tasks. A smaller model may be suitable for one and unreliable for another. The useful comparison is the quality and cost of the completed task, including retries and review.
Treat routing as part of the product
A routing rule can be wrong even when both models work well on their intended jobs. Evaluate the cases sent to each path and identify when an apparently simple request contains an exception. Keep the routing decision visible in logs so mistakes can be investigated.
Do not confuse confidence with correctness
A model’s confident wording is not sufficient evidence for accepting its output. Use observable checks such as valid structure, required references and recognised task boundaries. Where those checks cannot establish enough confidence, provide an explicit route to a stronger process or a human reviewer.
Source & context
OpenAI · GPT-5.4 mini and nano, 17 March 2026This retrospective was written for the archive in September 2026. The linked primary source documents the announcement or event; the practical interpretation and proposed approach are Sansa’s editorial perspective. Public examples do not imply a client relationship. Product capabilities and guidance may have changed since the period discussed.
Another perspective · March 2026
A two-tier document workflow with visible escalation
Continue readingWorking through a similar question?
Talk it through with Sansa