GPT-5.4

GPT-5.4 is a frontier large language model released by OpenAI, designed for complex professional work such as coding, research, automation, document processing and agent-based workflows. It is optimized for real-world tasks that require working with spreadsheets, presentations, documents, software tools and long workflows.

Key characteristics of GPT-5.4 include:

• Large 1 million token context window
• Native computer-use capabilities
• Advanced tool calling and tool search
• Improved agent workflows
• Higher benchmark performance across knowledge work and coding
• Improved token efficiency compared to GPT-5.2

Evolution of GPT Models

The GPT model family has evolved gradually with each version improving reasoning, coding, and tool interaction capabilities.

GPT-5.4 merges the best parts of GPT-5.2 reasoning and GPT-5.3-Codex coding capabilities.

Context Window and Token Limits

Context window refers to the amount of text the model can process in a single request. GPT-5.4 introduces one of the largest context windows available in frontier AI models.

Large context windows allow the model to process:

• Large research documents
• Full code repositories
• Long conversation histories
• Large datasets

Key Technical Improvements in GPT-5.4

1. Planning-Based Reasoning

GPT-5.4 Thinking can present an initial plan before generating the final answer. Users can modify this plan mid-response.

• Better control over responses
• Reduced back-and-forth iterations
• More accurate outputs

2. Improved Knowledge Work Capabilities

GPT-5.4 shows major improvements in professional tasks such as:

• financial modeling
• presentation generation
• spreadsheet creation
• document analysis

On the GDPval benchmark, which evaluates tasks across 44 occupations, GPT-5.4 achieved the highest performance.

This benchmark measures how well models perform professional knowledge tasks across industries.

3. Improved Spreadsheet and Document Work

GPT-5.4 significantly improves performance on real business tasks.

These improvements make GPT-5.4 suitable for professional workflows involving:

• financial spreadsheets
• reports
• presentations
• business documents

4. Reduced Hallucinations

GPT-5.4 improves factual accuracy. Compared to GPT-5.2:

• 33% fewer incorrect claims
• 18% fewer responses containing errors

This improvement increases reliability for professional work.

Computer Use Capabilities

GPT-5.4 introduces native computer interaction capabilities, allowing agents to perform tasks directly on software interfaces. Capabilities include:

• reading screenshots
• mouse clicks
• keyboard commands
• browser automation

This enables automation of tasks such as:

• sending emails
• scheduling calendar events
• filling forms
• navigating software interfaces

Computer Use Benchmarks

Performance improvements are measured using the OSWorld-Verified benchmark, which evaluates the ability to operate computers.

GPT-5.4 surpasses human performance in this benchmark.

Vision and Document Understanding

GPT-5.4 improves visual perception and document understanding. Example benchmarks:

The model also supports high-resolution image inputs. Image limits:

• Up to 10.24 million pixels
• Maximum dimension 6000 pixels

This improves tasks such as:

• document parsing
• chart understanding
• UI interaction

Coding Performance

GPT-5.4 integrates the coding strengths of GPT-5.3-Codex. Key improvements include:

• better code generation
• improved debugging
• faster execution
• lower latency

Benchmark comparison:

GPT-5.4 also performs well on Terminal-Bench 2.0, scoring 75.1%.

Tool Use Improvements

GPT-5.4 improves how models interact with external tools and APIs.

1. Tool Search

Previously, all tool definitions had to be included in the prompt. GPT-5.4 introduces tool search, which allows the model to fetch tool definitions only when needed.

• reduced token usage
• faster responses
• better scaling with large tool ecosystems

Example evaluation results:

This results in 47% token savings.

2. Agentic Tool Calling

GPT-5.4 also improves decision making when choosing tools. Example benchmark:

Toolathlon measures how well AI agents use tools to complete multi-step tasks.

3. Web Search Performance

GPT-5.4 improves persistent web search capabilities. Benchmark:

This benchmark evaluates the ability to find difficult information across multiple web pages.

Academic and Reasoning Benchmarks

GPT-5.4 shows strong improvements on reasoning benchmarks.

These benchmarks measure advanced reasoning ability.

Long Context Benchmarks

GPT-5.4 supports extremely long input contexts.

Safety and Deployment

GPT-5.4 is classified under OpenAI’s High Cyber Capability category. Security measures include:

• monitoring systems
• trusted access controls
• request-level blocking for high-risk requests
• improved safety classifiers

OpenAI also continues research on Chain-of-Thought monitoring to detect potential misuse.