GPT-5.6 Terra (medium) vs GPT-5 mini (high): The Ultimate Performance & Pricing Comparison

GPT-5.6 Terra (medium) is the stronger measured choice for demanding development work, while GPT-5 mini (high) is the much cheaper option. The available data shows Terra leading the coding index at 64.7 versus 15.6 and the intelligence index at 45.6 versus 25.3. GPT-5 mini (high) is priced at $0.6875 per 1M blended tokens, compared with $4.500000000000001 for Terra.\n\nThat headline is useful, but it is not a complete production recommendation. OpenAI’s current model directory documents the gpt-5.6-terra family alias as a Frontier model intended to balance intelligence and cost, but it does not provide a dedicated entry for gpt-5-6-terra-medium. The same directory does not independently list gpt-5-mini or confirm that high is a model identifier. See the OpenAI model directory.\n\nThe practical decision is therefore asymmetric. Terra has stronger measured coding and intelligence results, plus a reported median output speed of 119.568 tokens per second. Mini has a substantial price advantage and a math index of 90.7, but the comparison does not provide a corresponding Terra math result. Developers should treat Terra as the safer quality-first candidate and Mini as the cost-first candidate, subject to endpoint and model-ID verification.

GPT-5.6 Terra (medium) offers the clearer quality case, but GPT-5 mini (high) offers the clearer economic case.\n\n| Decision factor | Better evidence | What it means | |---|---|---| | Coding | GPT-5.6 Terra (medium) | The coding index is 64.7 versus 15.6, a large measured gap. | | General intelligence | GPT-5.6 Terra (medium) | The intelligence index is 45.6 versus 25.3. | | Math | GPT-5 mini (high) | Mini records 90.7, while no Terra math value is supplied. | | Blended cost | GPT-5 mini (high) | Mini costs $0.6875 versus $4.500000000000001 per 1M blended tokens. | | Output speed | GPT-5.6 Terra (medium) | Terra reports 119.568 median output tokens per second; Mini has no supplied value. | | Latency | Tie | Both models report 0.3 seconds. | \nThe evidence does not establish that Terra is better for every developer task. It establishes a strong advantage on the supplied coding and intelligence indices, while leaving the math comparison incomplete. It also does not establish whether either displayed name maps cleanly to a currently callable API model.\n\nOpenAI’s public documentation says that current models support text and image input, text output, multilingual capabilities, the Responses API, and official client SDKs, but it does not clearly assign every capability to these specific variants. The OpenAI model directory is therefore useful for family-level positioning, not for validating every variant-level assumption.\n\nData provided by https://artificialanalysis.ai/

GPT-5.6 Terra (medium) is the stronger candidate for code generation, code transformation, and multi-step engineering tasks based on the supplied evaluations.\n\nThe most important result is not that Terra wins several rows. It is the size and relevance of the coding difference: Terra scores 64.7 while Mini scores 15.6. For a developer workflow, that gap supports testing Terra first for repository-level changes, debugging, refactoring, and tasks where the model must preserve constraints across several files. The score does not prove that every generated patch will compile or pass tests. It does indicate that Mini should not be assumed to provide equivalent coding quality simply because both names belong to the same model family.\n\nTerra also leads the intelligence index, 45.6 to 25.3. That result strengthens the case for Terra when a request combines planning, code understanding, and judgment. It does not reveal which subskills drive the difference. The supplied material contains no official benchmark methodology, task distribution, or model-specific reliability report for either variant.\n\nMini’s math index is 90.7, but Terra has no supplied math value. That prevents a complete claim about mathematical superiority. A developer building a math-heavy product should run a task-specific evaluation instead of transferring Terra’s coding lead into an unsupported math conclusion.\n\nSpeed evidence is also one-sided. Terra reports 119.568 median output tokens per second, while Mini has no value in the data snapshot. Both models report 0.3 seconds of latency. That combination suggests Terra may stream long answers quickly, but it cannot establish that Terra delivers a faster end-to-end user experience in every application. Tool calls, prompt size, retries, and output length remain unmeasured here.

GPT-5 mini (high) is the clear price winner, but GPT-5.6 Terra (medium) can still be the lower-cost choice for workflows where failure creates rework.\n\nMini’s blended price is $0.6875 per 1M tokens, compared with $4.500000000000001 for Terra. Its input price is $0.25 and its output price is $2, while Terra is priced at $2 for input and $12 for output in the supplied snapshot. For high-volume classification, lightweight transformations, routing, and other tasks with modest quality requirements, Mini’s economics are difficult to ignore.\n\nThe price gap changes meaning when a response must be corrected by another model, reviewed by an engineer, or regenerated after a failed tool action. The data does not quantify those failure rates, so no break-even workload can be calculated responsibly. The coding index gap does provide a reason to test Terra for expensive engineering tasks, especially when a wrong patch consumes more time than the token bill.\n\nOpenAI’s pricing page lists Terra under standard, Batch, Flex, and Fast mode pricing, with short-context and long-context tiers. It does not list GPT-5 mini or confirm a current price for the displayed Mini variant. See the OpenAI API pricing page. That creates an operational risk: Mini may look cheaper in the supplied dataset, but the public page does not currently verify a callable, stable price for gpt-5-mini.\n\nTerra pricing also depends on processing mode and context tier. The pricing page does not publish the exact token boundary for short versus long context. Developers should measure actual request mix before selecting a billing mode, and should verify whether data residency adds the stated 10% surcharge for an eligible deployment.

GPT-5.6 Terra (medium) should be the default trial for quality-sensitive software development, while GPT-5 mini (high) should be the default trial for cost-sensitive high-volume work.\n\nChoose Terra when the model must modify an existing codebase, reason across constraints, explain a non-obvious failure, or produce an answer that is expensive to review. The supplied coding and intelligence indices support that choice. Terra’s reported 119.568 median output tokens per second also reduces concern that stronger reasoning necessarily means visibly slow streaming, although the comparison does not provide equivalent Mini speed data.\n\nChoose Mini when request volume dominates, individual responses are short, the task is easy to validate, or the system already has deterministic checks. Its $0.6875 blended price makes it suitable for broad first-pass processing. Its math index of 90.7 is a reason to test it for arithmetic or structured mathematical tasks, but the missing Terra math result means the comparison cannot identify a winner for that category.\n\nA staged architecture is plausible: use Mini for routine requests and escalate uncertain or high-impact cases to Terra. That approach is a design recommendation, not a measured result. The brief contains no routing accuracy, failure-rate, or escalation-cost data.\n\nBefore production adoption, verify the exact API model ID, current availability, supported parameters, context limits, output limits, and tool behavior. OpenAI’s model documentation does not independently confirm the gpt-5-mini entry or the high label. The pricing documentation does not currently list Mini pricing. Those documentation gaps matter as much as the benchmark gap.

GPT-5.6 Terra (medium) is the safer initial choice for developers who prioritize measured coding quality over token cost. The supplied coding index is 64.7 versus 15.6, but the official documentation does not provide a dedicated variant-level capability specification.\n\nGPT-5 mini (high) is the better initial choice for developers whose main constraint is request cost. Its blended price is $0.6875 per 1M tokens, yet the public pricing page does not independently verify that price for a currently listed gpt-5-mini model.\n\nNeither model can be declared the overall winner for math from the supplied evidence. GPT-5 mini (high) has a math index of 90.7, while the dataset provides no Terra math score.\n\nDevelopers should not assume high is a confirmed API model variant. The research brief found no official page connecting “GPT-5 mini (high)” to a model ID or confirming whether high represents a reasoning parameter.\n\nThe comparison does not establish which model has a larger context window. Both context-window values are null in the data snapshot, and OpenAI’s public pages do not provide the missing variant-specific limits.\n\nTerra has the stronger available speed evidence, but not a proven end-to-end latency advantage. Terra reports 119.568 median output tokens per second, while both models report 0.3 seconds of latency and Mini has no supplied output-speed value.