GPT-5.4

A General-Purpose Frontier Model for Real Work

OpenAI introduced GPT-5.4 as a model for professional work across the API and Codex. Its role is deliberately broad: developers can use it to analyze complex information, build production software, work with large collections of documents, and automate workflows that require several model and tool interactions.

Compared with earlier GPT-5 models, GPT-5.4 improved coding, instruction following, document understanding, image perception, tool use, and long-running task execution. OpenAI also emphasizes better token efficiency across tool-heavy workflows, where fewer unnecessary calls can matter as much as the price of each individual token.

GPT-5.4 was also an important architectural step in the model family. It introduced built-in computer use to the mainline model series, native compaction for longer agent trajectories, and Tool Search for large tool ecosystems.

• General-purpose frontier model for complex professional work.
• Strong software-engineering and code-heavy workflow support.
• Built for document-heavy and multi-step agentic tasks.
• First mainline GPT model with built-in computer-use support.
• Supports Tool Search and native compaction for long-running workflows.

Evaluate End-to-End Workflow Cost

For GPT-5.4, cost depends heavily on workflow shape. A simple single-turn request may be dominated by output tokens, while an agentic or document-heavy task can accumulate substantial input across source files, tool results, and conversation history.

Prompt caching can reduce the price of repeated stable prefixes. GPT-5.4 belongs to OpenAI's pre-GPT-5.6 caching generation: caching is automatic for eligible prefixes, a stable prompt_cache_key can help optimize routing, and there is no additional cache-write charge.

Agentic efficiency matters as well. OpenAI reports improvements in end-to-end performance and token efficiency on tool-heavy workflows compared with earlier models. Fewer unnecessary tool calls and shorter trajectories can reduce total workflow cost even when headline token rates stay fixed.

Very long prompts need separate budgeting. Once input exceeds 272K tokens, OpenAI applies higher long-context rates to the full session.

• $2.50 per 1M standard input tokens.
• $0.25 per 1M cached input tokens.
• $15.00 per 1M output tokens.
• No additional prompt-cache write charge.
• Tool-specific operations can introduce separate fees.

Start with the Lowest Effort That Meets the Quality Bar

The default none setting makes GPT-5.4 flexible for applications that want strong general-purpose behavior without automatically paying for deeper reasoning on every request.

More difficult workloads can move upward through low, medium, high, or xhigh. OpenAI's migration guidance recommends medium or high reasoning for workloads that previously depended on dedicated reasoning models such as o3, while tasks coming from GPT-4.1 can often begin at none.

This creates a practical evaluation strategy: keep a representative test set, establish the quality baseline at none, then increase reasoning only where the workload demonstrates a measurable benefit.

For professional applications, this is particularly useful because coding, document processing, and business analysis do not all need the same inference depth.