GPT-5.4 mini (xhigh)

GPT-5.4 mini (xhigh) is a mid-pack general model with a notably stronger coding position than its overall intelligence position. It ranks 63 of 578 on the Artificial Analysis Intelligence Index, while it ranks 56 of 202 on the Artificial Analysis Coding Index. That combination makes GPT-5.4 mini (xhigh) more interesting for software work than a broad leaderboard position alone suggests.

OpenAI lists gpt-5.4-mini as a model available through the Responses API and official Client SDKs, and its general model guidance describes text and image inputs, text outputs, and multilingual capability. The guidance does not clearly define the specific boundaries of this model. See OpenAI Models.

Data provided by https://artificialanalysis.ai/.

This review is useful for developers deciding whether the model belongs in an application, coding assistant, evaluation pipeline, or batch workflow. The evidence supports a practical value case, but it does not support confident claims about long-context behavior, maximum output length, or distinctive failure modes.

GPT-5.4 mini (xhigh) looks like a reasonable coding-oriented default when OpenAI compatibility matters and the workload does not require proven frontier performance. Its coding rank is 56 of 202, placing it in a stronger relative position than its general intelligence rank of 63 of 578. The result is good enough to justify testing, but not strong enough to remove task-specific validation.

The closest models provide a useful reference frame without changing the main conclusion:

Model	Practical trade-off against GPT-5.4 mini (xhigh)
Qwen3.6 Max Preview	Similar intelligence score, with a higher blended price of $2.925 per 1M tokens
GPT-5.2 Codex (xhigh)	Similar intelligence score, but a higher blended price of $4.8125 per 1M tokens
GLM-5.1 (Reasoning)	Slightly higher intelligence score and a lower coding score of 55.8
Grok Build 0.1 0616	Lower blended price of $1.25 per 1M tokens, but a lower coding score of 51.5
Inkling Small	Much lower blended price of $0.525 per 1M tokens, but a lower coding score of 52.9

OpenAI still lists gpt-5.4-mini in its official pricing catalogue. The same page does not clarify access for every account, region, or API endpoint. See OpenAI Pricing.

GPT-5.4 mini (xhigh) deserves developer attention because its coding rank is stronger than its overall intelligence rank. A position of 56 of 202 suggests that coding is a relative strength within the available evidence. It does not prove that the model will outperform a specialist on repository-scale work, but it supports trials for code generation, debugging, refactoring, and structured implementation tasks.

The practical implication is selection by workload. A team building a coding assistant should test repository navigation, patch accuracy, test repair, and instruction following separately. The benchmark does not reveal which of those tasks drives the coding result. It also does not reveal whether the xhigh label corresponds to a separate model or a reasoning configuration.

GPT-5.4 mini (xhigh) has a measured first-token latency of 0.3 seconds, which supports interactive use where users need an early response. Output throughput is not reported in the data brief, so the evidence cannot establish how quickly a long answer, code patch, or multi-step trace will finish. First-token responsiveness and total completion time should therefore be measured separately in production-like tests.

OpenAI’s model documentation confirms the model family and general input and output modalities, but it does not provide model-specific benchmark scores, context limits, or maximum output length. See OpenAI Models.

GPT-5.4 mini (xhigh) cannot yet be treated as a fully characterized production choice because key operational limits are not documented in the available evidence. The context window, maximum output length, supported API parameters, and model-specific official benchmarks were not found.

That gap matters most for developers handling large repositories, long conversations, document-heavy agents, or generated patches that may exceed ordinary response sizes. A model can look attractive on a coding index and still require application-level truncation, retrieval, or task decomposition. The available material does not show whether those precautions are necessary here.

The public evidence also does not identify reliable failure patterns. No dependable community reports were found that establish recurring coding mistakes, speed complaints, or model-specific quirks. That absence is not evidence of reliability. It means teams should create their own acceptance set before routing important workloads to the model.

The xhigh wording creates another naming risk. The official pricing page lists gpt-5.4-mini as the stable model identifier, but does not list xhigh as a separate priced model ID or explain the mapping. See OpenAI Pricing.

GPT-5.4 mini (xhigh) offers a strong cost-to-coding-position argument, but the value depends on how much output the application generates. Its blended price is $1.6875 per 1M tokens, with standard input priced at $0.75 and standard output priced at $4.50 per 1M tokens. The output rate makes verbose agents, repeated retries, and large generated patches more important to monitor than prompt volume alone.

The price is lower than the closest listed alternatives Qwen3.6 Max Preview and GPT-5.2 Codex (xhigh), while it is higher than Grok Build 0.1 0616 and Inkling Small. GPT-5.4 mini (xhigh) therefore occupies a middle position: it may reduce risk compared with cheaper models, but it should not be assumed to minimize spend.

Batch and Flex pricing are listed at $0.375 per 1M input tokens, $0.0375 per 1M cached input tokens, and $2.25 per 1M output tokens. Fast mode is listed at $1.50 per 1M input tokens and $9.00 per 1M output tokens. These modes can change the cost decision, but the suitable mode depends on workflow urgency and service availability.

The official pricing page does not list a long-context price. That omission does not prove that long context is unsupported, but it prevents a confident cost forecast for document-heavy applications. See OpenAI Pricing.

GPT-5.4 mini (xhigh) is a good candidate for coding assistants and developer automation when low first-token latency, OpenAI tooling, and moderate blended cost matter together. Its coding position of 56 of 202 is the strongest positive signal in the data brief. Developers should still gate adoption on task-level tests because the evidence does not explain the benchmark’s composition or the model’s failure behavior.

Use GPT-5.4 mini (xhigh) first when:

• the product already depends on OpenAI’s Responses API or official Client SDKs;
• coding tasks are more important than broad general reasoning;
• interactive users benefit from a 0.3-second first-token latency;
• the team can measure output volume, retries, and patch correctness;
• the application can tolerate undocumented context and output limits until verified.

Choose another model, or keep a fallback, when:

• the workload requires a documented context window or maximum output length;
• total generation speed is a hard requirement, because output tokens per second are not reported;
• the application needs known failure modes backed by public testing;
• the lowest possible token cost matters more than the coding rank;
• the xhigh configuration must be treated as a stable, independently documented model identity.

OpenAI’s public model catalogue shows that the product line includes newer GPT-5.6 models, but the available material does not state that gpt-5.4-mini is deprecated, replaced, or discontinued. Developers should verify availability before committing to a long-lived integration. See OpenAI Models.

GPT-5.4 mini (xhigh) is best evaluated as a coding-focused general model with incomplete public documentation. The questions below address the decisions that the available benchmark and research evidence can support.