GPT-5 (medium) vs GPT-5 mini (medium): The Ultimate Performance & Pricing Comparison

GPT-5 (medium) is the stronger measured option, while GPT-5 mini (medium) is the safer cost choice for high-volume applications. Artificial Analysis reports an Intelligence Index of 33.7 for GPT-5 (medium) and 30.9 for GPT-5 mini (medium), with Math Index scores of 91.7 and 85 respectively. The same data reports 0.3 seconds latency for each model, so the available evidence does not identify a latency winner.\n\nThe pricing gap is substantial. GPT-5 (medium) costs $3.4375 per 1M blended tokens, compared with $0.6875 for GPT-5 mini (medium). Input pricing is $1.25 versus $0.25, while output pricing is $10 versus $2.\n\nThe central selection problem is not simply quality versus price. The official OpenAI model directory does not list either requested model name, and the official pricing page does not list either configuration. That creates an availability and lifecycle risk that benchmark data alone cannot resolve. The current official documentation describes capabilities for the latest models generally, but it does not confirm that those statements apply to either requested configuration. See the OpenAI Models documentation and OpenAI API Pricing.\n\nData provided by Artificial Analysis.

GPT-5 (medium) leads the available quality evidence, but GPT-5 mini (medium) offers the more compelling default for workloads where token volume dominates engineering value.\n\n| Decision factor | GPT-5 (medium) | GPT-5 mini (medium) | Practical reading |\n|---|---:|---:|---|\n| Artificial Analysis Intelligence Index | 33.7 | 30.9 | GPT-5 (medium) has the higher measured general capability score |\n| Artificial Analysis Math Index | 91.7 | 85 | GPT-5 (medium) has the clearer advantage on the reported math evaluation |\n| Latency | 0.3 seconds | 0.3 seconds | The available snapshot shows a tie |\n| Blended price per 1M tokens | $3.4375 | $0.6875 | GPT-5 mini (medium) is the lower-cost option |\n| Input price per 1M tokens | $1.25 | $0.25 | Input-heavy workloads favor GPT-5 mini (medium) |\n| Output price per 1M tokens | $10 | $2 | Verbose generation makes the price difference more important |\n\nFor a developer building a first version, GPT-5 mini (medium) is the rational starting point when the task can tolerate some quality loss and the application processes substantial traffic. GPT-5 (medium) becomes more attractive when answer accuracy, mathematical reliability, or difficult reasoning has a direct business cost.\n\nThat recommendation remains conditional. Neither the official model directory nor the official pricing page confirms the requested model names, stable aliases, context limits, output limits, or current API availability. The research brief also found no reliable community reports for coding behavior, speed perception, or recurring failure patterns. The evidence supports a measured capability and cost comparison, but not a complete production-readiness judgment.\n\nThe official documentation describes text and image input, text output, multilingual ability, vision, Responses API access, and official SDK access for latest models generally. It does not establish that either requested model receives every listed capability. The relevant source is the OpenAI Models documentation.

GPT-5 (medium) has the stronger reported performance profile, especially where mathematical reliability and difficult reasoning affect downstream decisions.\n\nThe Artificial Analysis Intelligence Index places GPT-5 (medium) at 33.7 and GPT-5 mini (medium) at 30.9. The Math Index separates them more clearly, at 91.7 and 85. These results support choosing GPT-5 (medium) for tasks where a wrong answer creates manual review, failed automation, or user distrust. They do not prove that GPT-5 (medium) wins every coding, tool-use, writing, or retrieval task. The brief provides no task-level benchmark breakdown that would justify that broader claim.\n\nThe practical meaning of the difference depends on the application’s error budget. A small quality gap may be acceptable for classification, routing, extraction with validation, or short drafts. The same gap may matter more for numerical analysis, complex planning, code generation, or responses that users act on without review. Developers should test representative prompts rather than treating an index score as a universal ranking.\n\nLatency does not currently resolve the choice. Both models are reported at 0.3 seconds, and median output speed is unavailable for both. The snapshot therefore supports a latency tie, but it cannot establish streaming behavior, throughput under concurrency, time to first token, or long-response performance. Those missing measurements are important for interactive developer tools.\n\nThe official OpenAI model directory does not provide model-specific context windows, maximum output lengths, API parameters, or benchmark results for either requested configuration. That omission limits what can be inferred about long-context coding, large repository analysis, and structured output reliability. The source is the OpenAI Models documentation.

GPT-5 mini (medium) is the clear price winner, but its lower token rate does not automatically produce the lower total engineering cost.\n\nGPT-5 mini (medium) is listed at $0.6875 per 1M blended tokens, versus $3.4375 for GPT-5 (medium). Its input price is $0.25 versus $1.25, and its output price is $2 versus $10. Those differences favor GPT-5 mini (medium) for high-volume prompts, repeated summarization, routing, extraction, and other workloads where each request has limited consequence.\n\nThe cost conclusion can change if the cheaper model produces more retries, longer prompts, more validation calls, or more human review. A model that misses important constraints may require a second model call or a fallback path. A model that generates less useful code may increase developer time even when API spending falls. The data brief does not measure retry rates, output length, correction rates, or human review costs, so it cannot determine total cost of ownership.\n\nGPT-5 (medium) may be economically justified when its higher measured Math Index of 91.7 reduces costly errors compared with 85 for GPT-5 mini (medium). That claim remains a hypothesis for a specific application, not a measured savings result. Teams should compare complete workflows, including validation and fallback behavior, rather than comparing token prices alone.\n\nThe official pricing page does not list either requested model name or provide a confirmed current price for either configuration. The prices in this comparison come from the supplied Artificial Analysis snapshot, not from current OpenAI pricing documentation. See OpenAI API Pricing and Artificial Analysis.

GPT-5 mini (medium) is the better first deployment candidate for cost-sensitive, high-volume workflows, while GPT-5 (medium) deserves escalation for quality-critical tasks.\n\nChoose GPT-5 mini (medium) when the application processes many routine requests, has strong validation, or can accept occasional quality differences. Its $0.6875 blended price and $2 output price make it better suited to applications where token consumption is a primary constraint. This includes internal assistants, document triage, lightweight extraction, support classification, and early product experiments. The available evidence does not confirm its stable API alias or current direct availability, so deployment should begin only after an endpoint check.\n\nChoose GPT-5 (medium) when errors carry a material cost or when the task requires stronger reasoning. Its Intelligence Index of 33.7 exceeds 30.9, and its Math Index of 91.7 exceeds 85. Those results make it the better candidate for numerical reasoning, difficult planning, code review, and automation that needs fewer corrections. They still do not replace a task-specific evaluation.\n\nUse a routing strategy only if the product can measure failure outcomes. Start with GPT-5 mini (medium) for low-risk requests, then escalate selected cases to GPT-5 (medium) when validation detects ambiguity or unacceptable uncertainty. The brief contains no evidence about confidence calibration, tool use, coding style, or failure triggers, so the routing rule must come from application telemetry rather than assumed model behavior.\n\nThe largest unresolved issue is product status. OpenAI’s current model directory does not list either requested name, and it does not explain whether either model was replaced, renamed, or made unavailable. The OpenAI Models documentation should be checked before implementation, while the OpenAI API Pricing page should be checked before budgeting.

GPT-5 (medium) is the quality-oriented choice, while GPT-5 mini (medium) is the cost-oriented choice, based on the supplied benchmark and pricing snapshot.\n\nThe available evidence does not answer whether either requested configuration is currently callable through a stable OpenAI API alias. The official model directory omits both names, and the research brief found no confirmed migration or replacement statement.\n\nDevelopers should treat the Artificial Analysis figures as directional selection evidence. The figures identify relative benchmark and pricing differences, but they do not cover context limits, output limits, throughput, streaming speed, tool behavior, coding outcomes, or production failure rates.\n\nA sound evaluation should replay representative application prompts, record correction and fallback behavior, and include the cost of review. That process is necessary because the available materials do not directly measure total workflow quality or total workflow cost.