GPT-5.4 mini

OpenAI identifies GPT-5.4 mini as a 'distilled' variant of the flagship GPT-5.4 model, but provides no technical documentation regarding the distillation process, architectural modifications, or the specific transformer configuration. While it is described as a 'unified architecture' that merges reasoning and coding capabilities, there is zero public information on layers, attention mechanisms, or pre-training methodology beyond vague marketing descriptions of efficiency.

OpenAI provides no specific breakdown of the training data for GPT-5.4 mini. Documentation only mentions a knowledge cutoff of August 31, 2025, and claims the model is 'factually grounded' with 33% fewer false claims than GPT-5.2. No information is provided regarding the ratio of web data, code, or books, nor are there details on filtering, cleaning, or the use of synthetic data during the distillation process.

The model utilizes the 'o200k_base' tokenizer with a vocabulary size of approximately 200,000 tokens. While the tokenizer's behavior can be observed via the API and third-party tools, OpenAI has not released a formal technical paper documenting the specific BPE configuration or training alignment for the 5.4 series. This lack of official documentation for the specific version limits full verification of its multilingual efficiency.

The parameter count for GPT-5.4 mini is completely undisclosed. While it is marketed as a 'small' and 'efficient' model that runs 2x faster than its predecessor, OpenAI provides no data on total or active parameters. There is no clarification on whether the architecture is dense or utilizes sparse Mixture-of-Experts (MoE) techniques, making it impossible to verify its density claims.

OpenAI provides no information regarding the compute resources used to train or distill GPT-5.4 mini. There are no disclosures on GPU/TPU hours, hardware specifications, training duration, or carbon footprint. The company explicitly omits these details in official announcements, citing competitive reasons.

While OpenAI provides specific scores for benchmarks like SWE-Bench Pro (54.4%) and OSWorld-Verified (72.1%), the evaluation code, exact prompts, and few-shot examples are not public. Third-party analysis notes that comparisons are often made against older models (GPT-5.2) rather than the most recent predecessors, and the lack of a technical paper prevents independent verification of the claimed results.

The model consistently identifies itself as GPT-5.4 mini across API calls and system prompts. It maintains clear versioning (gpt-5.4-mini) and correctly identifies its capabilities, such as its 400K context window and multimodal support. There are no documented instances of the model claiming to be a competitor's product or misrepresenting its identity.

The model is governed by a restrictive proprietary license. While the terms for API usage and commercial pricing ($0.75/M input, $4.50/M output) are clearly stated, the underlying weights and code are not accessible. The license is subject to OpenAI's standard Terms of Service, which can change without notice, and provides no transparency for derivative works or offline use.

As a closed-source API model, there is no documentation on the VRAM or hardware requirements for local deployment. OpenAI provides latency and throughput metrics (e.g., ~180-190 tokens/sec), but these are service-side performance figures rather than hardware footprint data. There is no guidance on quantization tradeoffs or memory scaling for the 400K context window.

OpenAI uses semantic-like versioning and provides 'snapshots' to allow developers to lock in specific model versions for consistency. However, the company has a history of 'silent updates' and performance drift that is not always captured in public changelogs. While version history is accessible via the API, the lack of detailed technical changelogs for architectural or weight updates limits transparency.