GPT-5.4 Test: Human-Level AI for Professional Tasks

OpenAI has launched GPT-5.4, positioned as a significant advancement in AI models, combining reasoning, coding, and agentic workflows. This release follows a series of incremental updates (5.1, 5.2, 5.3) but carries higher expectations due to its theoretical connection to OpenAI's 'Code Red' initiative from December. Early reactions from testers like Ben Hilac indicate that GPT-5.4 is a noteworthy model, suggesting it's the first in a long time worth trying.

OpenAI frames GPT-5.4 as a model designed for professional work, contrasting it with GPT-5.3 Instant, which focused on speed and personality for personal use cases. The model incorporates industry-leading coding capabilities from GPT-5.3 Codecs and improves its functionality across tools, software environments, and professional tasks like spreadsheets and presentations. This focus aims to deliver accurate, effective, and efficient results with less back-and-forth.

A significant feature of GPT-5.4 is its 1 million token context window, which enhances its ability to handle tasks requiring extensive thought. Early testers, such as Brendan Foody, CEO of Merkore, praise GPT-5.4 as the best model they've tried, excelling in creating long-horizon deliverables like slide decks and financial models. It achieves top performance while running faster and at a lower cost than competitive frontier models.

Efficiency is a core theme, with GPT-5.4 being described as OpenAI's most token-efficient reasoning model, using significantly fewer tokens and offering faster speeds compared to GPT-5.2. The Codecs' fast mode delivers up to 1.5x faster token velocity. Tool search capabilities have also been optimized; instead of including all tool definitions upfront, GPT-5.4 uses a lightweight list and looks up definitions only when needed, reducing token usage by 47% on evaluated tasks while maintaining accuracy.

GPT-5.4 demonstrates remarkable improvements in computer use, a critical capability in the 'OpenClaw world.' Rahul Agaral notes that GPT-5.4 can use a computer better than a human, operating websites and software autonomously, issuing keyboard/mouse commands, and navigating desktop environments. On the OSWorld verified benchmark, it scored 75%, surpassing human-level performance (72.4%) and significantly outperforming GPT-5.2 (47.3%).

Jamie Cuff from PACE stress-tested GPT-5.4 on complex legacy insurance portals, finding a paradigm shift in AI's ability to navigate difficult UIs. Key improvements include vastly better click accuracy, even on crowded screens, and enhanced long trajectory reasoning, speed, and memory. This advancement shifts the automation bottleneck from 'can the model do it' to 'do you trust it enough to let it,' posing new challenges for deployment.

The GDP val benchmark, which measures performance on knowledge work across 44 occupations, shows GPT-5.4 achieving a win rate of 69.2% to 70.8% against industry professionals. When ties are included, this figure rises to 82-83%. Ethan Mollik calculates that for a 7-hour task, GPT-5.4 could save an average of 4 hours and 38 minutes, even accounting for failures and result checking.

OpenAI is aggressively targeting specific industries; COO Brad LCAP highlighted GPT-5.4's improvements for finance, including enhanced financial modeling and analysis, direct Excel integration, and connections to platforms like Factivia and S&P Global. This specialized focus suggests a 'Codecs moment' for professional services, where the model's capabilities could significantly impact industry workflows.

Despite its powerful capabilities, GPT-5.4 has notable drawbacks in user experience. Testers frequently observed over-verbosity, with the model producing excessively long responses and repeating itself. This 'eagerness' to expand tasks beyond initial requests and its tendency to stay in planning mode rather than immediately executing tasks can create a significant cognitive burden for the user.

A major criticism is GPT-5.4's poor UI design taste. Ben Davis and Matt Schumer noted its 'hilariously bad' and 'tasteless' visual output, with one critique describing its designs as 'muddy gradient blobs' and 'dull and washed out.' This deficiency often necessitates using other tools, like Claude, for front-end design work. However, the updated Codecs CLI experience is praised for its reduced friction in the approval system and transparent interstitial updates during long-running tasks, improving the developer workflow.