GPT-5 mini (medium)

Instead of asking for open-ended prose, engineer your prompts to demand brevity and structure. This is the single most effective way to manage its high output cost.

• Instruct the model to respond with bullet points, a numbered list, or a short summary.
• For data processing, demand JSON output with a predefined schema. This is often far more token-efficient than a natural language explanation.
• Add constraints to your prompt, such as: "Explain this in three sentences," or "Provide only the final answer."

The model's affordable input pricing and large context window create an opportunity for batch processing. Consolidate multiple, smaller tasks into a single API call to reduce overhead and potentially improve consistency.

• Instead of asking ten questions in ten separate calls, combine them into one prompt and ask for a numbered list of answers.
• Feed a large batch of items (e.g., customer reviews) and ask for a single summary of sentiment or a JSON array of classifications.
• This approach is ideal for offline processing where latency is not a concern.

Never make an API call without setting a sensible `max_tokens` parameter. This acts as a crucial safety net to prevent the model from generating excessively long (and expensive) responses, especially if a prompt is unintentionally ambiguous.

• Calculate the expected length of a reasonable response for your use case and set the limit just above it.
• For classification or extraction tasks, the limit can often be set very low (e.g., under 100 tokens).
• Monitor for responses that are truncated by the limit, as this may indicate a need to adjust your prompt or the limit itself.

Reserve GPT-5 mini (medium) for the tasks where its intelligence is indispensable. For other parts of your workflow, use cheaper, faster models. This "model routing" or "cascade" approach optimizes for both cost and quality.

• Use a fast, inexpensive model for initial intent classification or to handle simple, repetitive queries.
• When a query is identified as complex or high-value, escalate it to GPT-5 mini (medium) for the final, authoritative answer.
• For drafting content, use a cheaper model to generate a first pass, then use GPT-5 mini (medium) for a final step of refinement and editing.

GPT-5 mini (medium) is a model from OpenAI positioned as a balance between the top-tier GPT-5 models and more efficient, smaller versions. It is characterized by very high intelligence, a large 400,000-token context window, and multimodal (text and image) input capabilities, but comes with a high output price and slower-than-average speed.

Compared to models with a similar blended price, GPT-5 mini (medium) is typically more intelligent and has a much larger context window. However, it is often slower and has a uniquely skewed cost structure, with its output tokens being significantly more expensive than its peers. The trade-off is elite reasoning for a premium generation cost.

This model excels at tasks that are input-heavy and require deep understanding. Ideal use cases include: in-depth analysis of long legal or financial documents, building sophisticated question-answering systems over large knowledge bases (RAG), extracting structured data from unstructured text, and powering expert-level chatbots where accuracy is more critical than speed.

This pricing strategy reflects the underlying computational costs. Processing input tokens (reading) is generally less computationally intensive than generating new tokens (writing), especially for a highly complex model. The high output price subsidizes the model's advanced generative capabilities, while the low input price encourages users to leverage its large context window.

While the model technically supports a 400k token context, using the full window has practical implications. API calls with very large inputs will have higher latency and can be expensive despite the low per-token input cost (e.g., 350k input tokens still cost ~$0.09). It is most effective when the task genuinely requires access to that entire block of information at once, a concept often referred to as "needle-in-a-haystack" evaluation.

The choice depends on your priorities. Microsoft Azure shows a slight performance advantage in our testing, with lower latency and higher throughput. However, the difference is marginal. Developers may prefer OpenAI for its straightforward API and ease of integration, while organizations already embedded in the Microsoft ecosystem may find Azure a more natural fit. As pricing is identical, the decision can be based on performance needs and platform preference.