GPT-5 Mini

Intelligence Designed for Repetition at Scale

GPT-5 Mini occupies a different role from flagship and Pro models.

Instead of spending premium inference on the hardest possible problem, Mini targets workloads that need useful reasoning repeatedly: structured processing, focused analysis, application logic, extraction, classification, routing, concise generation, and other tasks with a clear definition of success.

That distinction matters in production. A small cost difference per request becomes substantial when an application executes millions of requests or uses a model repeatedly inside a workflow.

OpenAI describes GPT-5 Mini as strong intelligence for cost-sensitive, low-latency, high-volume workloads and specifically notes that it performs well on well-defined tasks with precise prompts.

Its 400K context window also means "Mini" describes the model's efficiency tier rather than a tiny context budget. Applications can still provide substantial documents, histories, retrieved evidence, and instructions.

GPT-5 Mini is now deprecated. For most new workloads in the same latency-and-volume category, OpenAI recommends GPT-5.6 Terra.

• Lower-cost member of the GPT-5 family.
• Optimized for well-defined and precisely prompted tasks.
• Designed around cost-sensitive, high-volume production.
• 400K context despite its efficiency positioning.
• Deprecated in favor of newer efficient models.

Small Unit Costs Matter Most at Large Request Volumes

GPT-5 Mini's pricing makes the most sense when viewed through aggregate workload economics.

A single request may cost fractions of a cent. Multiply that request across a large customer base, batch pipeline, document collection, or multi-step workflow and model selection becomes a meaningful infrastructure decision.

Cached input is priced at one tenth of standard input. Repeated prompt prefixes, stable instructions, or recurring context can therefore materially reduce input cost when caching applies.

Output remains more expensive than input, so concise schemas and bounded generation are useful for both latency and spend.

For production evaluation, measure cost per completed business operation. A cheaper model that requires retries or frequent escalation may not be cheaper at the workflow level.

• $0.25 per 1M input tokens.
• $0.025 per 1M cached input tokens.
• $2.00 per 1M output tokens.
• Cached input is 10% of standard input price.
• Evaluate effective cost after retries and fallback routing.

Reasoning Should Match the Value of the Subtask

The model's strongest production role is not simply "cheap text."

Reasoning support allows GPT-5 Mini to handle tasks that involve interpretation, rule application, multi-step decisions, or structured analysis rather than only surface-level completion.

The engineering question is whether the task benefits enough from reasoning to justify the additional inference work.

For high-volume systems, this should be measured empirically. A reasoning configuration that improves accuracy by a small amount may be highly valuable for one workflow and unnecessary for another.

Use representative evaluation sets instead of assuming one reasoning policy should cover extraction, classification, generation, and agent subtasks equally.