GPT-5 (medium)

Given the model's tendency to be verbose and the high cost of output tokens, controlling generation length is the most important cost-saving measure.

• Use `max_tokens`: Always set a reasonable `max_tokens` limit in your API calls to prevent runaway generation and cap the cost of any single request.
• Prompt Engineering: Explicitly instruct the model to be concise. Use phrases like "Be brief," "Answer in one sentence," or "Provide a bulleted list of 3 items."
• Structure the Output: Ask the model to return a structured format like JSON. This often results in less conversational, more compact output that is also easier to parse programmatically.

Your goal is to minimize expensive output tokens relative to cheap input tokens. Design your application logic around this principle.

• Favor 'Extraction' over 'Generation': Frame tasks as information extraction from a large context whenever possible. For example, instead of asking "Write a summary of this document," ask "Extract the key 3 findings from this document."
• Use Few-Shot Prompting for Formatting: If you need a specific output format, provide examples in the prompt (input tokens) rather than asking the model to generate a lengthy explanation of the format before providing the answer (output tokens).

The massive context window is a powerful feature, but using it inefficiently can be costly and slow, even with cheap input tokens.

• Don't Send Redundant Context: For session-based applications (like chatbots), implement a summarization strategy for the conversation history instead of passing the entire transcript with every turn.
• Batch Queries: If you need to ask multiple questions about the same large document, pass the document once and ask all questions in a single, structured API call to avoid repeatedly paying the input cost for the same context.
• Consider a RAG Pipeline: For very large knowledge bases, a traditional RAG approach (embedding and retrieving relevant chunks) may still be more cost-effective than stuffing hundreds of thousands of tokens into the context for every query.

While pricing is identical, the performance difference between OpenAI and Azure is significant.

• Default to Azure for Performance: If your application is user-facing or requires low latency, the speed and responsiveness of the Azure endpoint are worth the potentially more complex setup.
• Use OpenAI for Prototyping: The direct OpenAI API is perfect for quick experiments, internal tools, and applications where a few seconds of extra latency don't matter.

GPT-5 (medium) is a large language model from OpenAI. It is 'multimodal,' meaning it can process both text and images as input. It is distinguished by its high intelligence score, a very large 400,000-token context window, and strong performance, particularly when served via Microsoft Azure.

GPT-5 (medium) is a generational leap over GPT-4 Turbo. It demonstrates significantly higher intelligence and reasoning capabilities, as shown by its score of 66 on the Intelligence Index. It also features a much larger context window (400k vs. 128k) and is substantially faster, especially on optimized infrastructure like Azure.

The 'medium' designation likely refers to its position within the broader GPT-5 family of models. It suggests that OpenAI may also offer other versions, such as a smaller, faster 'light' model or a larger, even more powerful 'large' or 'ultra' model. 'Medium' implies a balance between cutting-edge capability and manageable operational cost and speed.

The performance difference stems from the underlying infrastructure. Microsoft has heavily invested in optimizing its Azure data centers specifically for serving large OpenAI models at scale. This includes specialized hardware, networking, and software configurations that allow them to achieve higher throughput and lower latency than the general-purpose OpenAI API infrastructure.

Not always. While incredibly powerful for analyzing large documents or maintaining long conversations, it comes with trade-offs. Filling the context window increases cost (a full 400k context costs $0.50 in input tokens) and can increase the model's response time. For many tasks that don't require that much context, using a smaller amount is more efficient and cost-effective.

GPT-5 (medium) excels at tasks that require a combination of deep reasoning, large context, and speed. Top use cases include: advanced Retrieval-Augmented Generation (RAG) over large document sets, complex legal or financial document analysis, multimodal applications that analyze charts or diagrams, and sophisticated, low-latency customer-facing chatbots that require extensive conversation history.