GPT-5.4: Everything You Need to Know

OpenAI's GPT 5.4: Key Features and Improvements

Overview of GPT 5.4

• OpenAI has released GPT 5.4, which features native computer use capabilities, allowing it to operate through user interfaces (UIs) and navigate desktops effectively.

• The model supports an experimental codec API with a context window of 1 million tokens, significantly increasing from the previous limit of 272,000 tokens.

Enhanced Interaction and Efficiency

• Users can now interrupt GPT 5.4 during its processing without needing to restart the task, enhancing workflow efficiency.

• Tool search functionality allows the model to look up tool definitions on demand, reducing token usage by approximately 47%.

Performance Benchmarks

• GPT 5.4 scores an impressive 83% on the GDP wall benchmark, indicating strong performance in knowledge work tasks like spreadsheets and presentations.

• Compared to its predecessor (GPT 52), there is a notable improvement of about 20% in performance at similar costs due to enhanced cost and token efficiency.

Comparison with Other Models

• Despite being a significant upgrade over GPT 52, on normal settings, GPT 5.4 lags behind Gemini 3.1 Pro but excels in extra high settings.

• The coding capabilities have improved considerably; however, when compared directly with the specialized Codex model (GPT53), improvements are less pronounced.

Reasoning Capabilities

• Medium reasoning efforts show greater accuracy increases than high reasoning efforts in benchmarks, suggesting a balance between speed and accuracy that benefits coding tasks.

• On design benchmarks, GPT 5.4 ranks ninth at medium settings—a substantial jump from previous versions—indicating improved UI design capabilities.

Significant Benchmark Insights

GDP Benchmark Analysis

• The GDP benchmark assesses models' effectiveness across various professional occupations contributing to US GDP; it reflects how well these models perform knowledge work.

• From GPT52 to GPT54, there is a remarkable increase of nearly 12% in correct responses on this challenging benchmark.

Release Cadence and Training Enhancements

• OpenAI has accelerated its release schedule significantly; three major releases occurred within four months since November for version one.

• Post-training improvements specifically target computer use capabilities using libraries like Playwright have led to better interaction with existing systems.

Coding Capabilities Comparison

Performance Metrics Against Previous Versions

• While there is substantial improvement over GPT52 regarding coding tasks, comparisons with Codex (GPT53) reveal only marginal gains—around a single percentage point improvement in OS world verified benchmarks.

Understanding Token Efficiency and Tool Search in AI Models

Reasoning Effort and Token Efficiency

• Setting the reasoning effort to medium results in a significant improvement compared to previous models (GPD 53), emphasizing the need for token efficiency and faster code generation.

• The introduction of tool search aims to enhance token efficiency by avoiding the pollution of context windows with unnecessary tool definitions, a problem noted in traditional agentic coding systems.

Tool Search Implementation

• OpenAI's new tool search feature loads only a limited number of tools into the context window, allowing for more relevant tool selection based on task requirements.

• Enabling tool search can reduce token usage by nearly half, showcasing its effectiveness in improving overall performance, particularly in agentic tool use.

Performance Improvements

• The model demonstrates enhanced capabilities in web searches due to improved computer use functionalities, indicating better performance across various tasks.

• Users can now steer the model during its processing by providing additional information, enhancing interaction and control over outputs.

Pricing Considerations

• Compared to GPT52, GPT54 is more expensive; however, its increased token efficiency may offset costs. The Pro version is significantly pricier but targets specialized scientific research rather than general users.