GPT-5.4

GPT-5.4 lists base rates of $2.50 per million input tokens and $15 per million output tokens. Estimate both input and output token volume and any additional charges for your workload. All amounts are USD per MTok; long-context bands apply to the full request, not just excess tokens. Standard <=272K input: input/output $2.50/$15.00, cached input $0.25, 30-minute write not quoted, 12-hour write not quoted; >272K: input/output $5.00/$22.50, cached input $0.50, 30-minute write not quoted, 12-hour write not quoted. Batch <=272K input: input/output $1.25/$7.50, cached input $0.13, 30-minute write not quoted, 12-hour write not quoted; >272K: input/output $2.50/$11.25, cached input $0.25, 30-minute write not quoted, 12-hour write not quoted. Flex <=272K input: input/output $1.25/$7.50, cached input $0.13, 30-minute write not quoted, 12-hour write not quoted; >272K: input/output $2.50/$11.25, cached input $0.25, 30-minute write not quoted, 12-hour write not quoted. Fast <=272K input: input/output $5.00/$30.00, cached input $0.50, 30-minute write not quoted, 12-hour write not quoted; >272K: input/output not quoted/not quoted, cached input not quoted, 30-minute write not quoted, 12-hour write not quoted. Regional processing and FedRAMP add 10% where the model and route are eligible; region eligibility is separate from a quoted tariff. Server-side tools and their tokens can add charges. The model card applies the >272K surcharge to the full session in Standard/Batch/Flex. Row captions say <272K while the table band definition says <=272K; confirm any boundary-sensitive billing. Cache-write cells are not quoted.

GPT-5.4 accepts about 1.05M tokens of context. That only matters if you routinely send very long documents, large codebases, or multi-turn histories that approach that limit.

GPT-5.4 is a balanced tier from OpenAI. It suits general production, agentic workflows, knowledge work. An older API model with a 1.05M context. Published API availability and specs do not establish adoption, support quality or superiority to newer models; compare your workload with current GPT-6 options.