GPT-5.4 is OpenAI’s latest frontier model, unifying the Codex and GPT lines into a single system. It features a 1M+ token context window (922K input, 128K output) with support for text and image inputs, enabling high-context reasoning, coding, and multimodal analysis within the same workflow.
The model delivers improved performance in coding, document understanding, tool use, and instruction following. It is designed as a strong default for both general-purpose tasks and software engineering, capable of generating production-quality code, synthesizing information across multiple sources, and executing complex multi-step workflows with fewer iterations and greater token efficiency.
Evaluations
16
across 14 benchmarks
Latency
494ms
Context length
1M
tokens
Cost
$2.50 · $15
input · output per 1M tokens
Key takeaways
GPT-5.4 is OpenAI’s latest frontier model, unifying the Codex and GPT lines. It is designed for high-context reasoning, coding, and multimodal analysis, supporting text and image inputs. It serves as a strong default for general-purpose tasks and software engineering.
The model utilizes a proprietary Mixture-of-Experts (MoE) architecture with a 1M+ token context window (922K input, 128K output). It is trained to generate production-quality code, synthesize information across multiple sources, and execute complex multi-step workflows.
GPT-5.4 demonstrates strong arithmetic and algebraic reasoning (94.6% accuracy on MATH-500) and adheres to complex output formatting. It also shows strong code generation capabilities with 95.7% accuracy on human evaluation tasks.
GPT-5.4 struggles with precision in numerical tasks, nuanced factual recall in biology/medicine, complex logical interpretation, and image-based analysis, often showing overconfidence in incorrect answers (10.67% accuracy on Humanity's Last Exam).
Minor discrepancies in passing coding tasks typically arise from subtle differences in interpreting problem details or expectations for exact algorithmic approach, rather than fundamental coding errors. The model failed all tests (0% accuracy) on the ARC-AGI-2 dataset.
