GPT-5.4 Pro

OpenAI identifies GPT-5.4 Pro as a 'unified' model that consolidates the reasoning of the GPT-5 series with the coding capabilities of GPT-5.3-Codex. However, the underlying architecture is described only as 'dense' in marketing materials, with no public technical report or documentation detailing the specific training methodology, architectural modifications, or pretraining procedures. The lack of a peer-reviewed paper or detailed technical specification makes the architectural provenance largely unverifiable.

There is no public disclosure of the training data sources, composition breakdown, or data collection methodology. Documentation only mentions a knowledge cutoff of August 31, 2025, and vague claims of being 'factually grounded' with a 33% reduction in errors. No information regarding filtering, cleaning, or the use of synthetic data is available, which is a significant transparency gap for a frontier model.

The tokenizer is accessible via the OpenAI API, and the model supports a standard 272K context window (expandable to 1.05M tokens). However, the specific vocabulary size, BPE configuration, and training alignment for the 5.4 version are not explicitly documented in any public technical report. While behavior can be observed through API usage, the underlying technical specifications remain opaque.

The parameter count for GPT-5.4 Pro is officially 'Unknown.' While it is marketed as a 'dense' architecture, there is no verifiable information regarding the total number of parameters or the architectural breakdown (e.g., attention vs. FFN). This lack of disclosure is a major transparency failure for a model positioned as a high-performance enterprise variant.

OpenAI has not disclosed any specific details regarding the GPU/TPU hours, hardware specifications, or energy consumption used to train GPT-5.4 Pro. There are no carbon footprint calculations or cost estimates provided in official documentation, relying instead on vague marketing claims of being the 'most efficient' frontier model.

While OpenAI provides specific scores for several benchmarks (e.g., 83.3% ARC-AGI-2, 89.3% BrowseComp), the evaluation code, exact prompts, and few-shot examples are not fully public. Some benchmarks, such as GDPval, are described as 'in-house,' making independent third-party verification and reproduction impossible. A -2 penalty was applied due to partial disclosure of contamination issues in predecessor models without clear mitigation details for this version.

The model consistently identifies itself as GPT-5.4 Pro and maintains clear versioning (5.2, 5.3, 5.4) within the API and ChatGPT interface. It is transparent about its capabilities regarding native computer use and extended context. However, it loses points for occasional confusion in complex agentic workflows where it may second-guess its own identity or output.

The model is released under a proprietary license with no open-source access to weights or source code. While commercial use is permitted via the API, the terms are governed by OpenAI's Terms of Service, which can change without notice. The lack of a clear, stable license for derivative works or weight inspection results in a low score.

As a closed-source API-based model, there is no public documentation regarding VRAM requirements or hardware footprints for local deployment. OpenAI provides pricing based on token usage but offers no guidance on quantization tradeoffs or memory scaling for the experimental 1M token window, leaving developers with no information on the actual compute resources required to run the model.

OpenAI uses semantic-like versioning and provides 'snapshot' features in the API to mitigate drift. However, there have been documented reports of 'stealth nerfs' and performance degradation (e.g., reduced thinking time) following plan changes, which are not reflected in official changelogs. A -3 penalty was applied for behavior changes occurring without corresponding version updates.