OpenAI GPT-5.3 & GPT-5.4: New Models Boost Performance & Context Window

0 comments

OpenAI Unveils GPT-5.4: A Leap Forward in AI Capabilities

OpenAI has released its next-generation language models, GPT-5.3 Instant and GPT-5.4, marking a significant evolution in artificial intelligence. While GPT-5.3 Instant focuses on speed and conversational flow for everyday use, GPT-5.4, available in Thinking and Pro versions, is designed for more complex reasoning and professional applications. OpenAI announced the updates on March 5, 2026.

GPT-5.3 Instant: Enhanced Conversational AI

GPT-5.3 Instant aims to deliver more accurate and contextually rich responses, reducing unnecessary caveats and overly cautious phrasing that can disrupt natural conversation. 9to5Mac reports that the model is designed to better integrate its own knowledge with current web information, providing context for breaking news and minimizing extraneous links.

Internal tests indicate a 26.8% lower rate of hallucinations in medical, legal and financial queries when using web search, and a 19.7% reduction when relying on internal knowledge. Currently, improvements are primarily focused on English language responses, with ongoing function to enhance other languages.

GPT-5.4: Powering Advanced Workflows

GPT-5.4 combines the coding prowess and computer skills of the GPT-5.3 Codex model with improved general responsiveness, web usage, and context maintenance. A key feature is the ability to interrupt and guide the model’s thinking process, allowing users to refine outputs mid-generation without restarting. TechRadar highlights this as a major step forward for complex projects.

OpenAI highlights six key improvements in GPT-5.4: coding, document understanding, tool use, instruction following, image perception and multimodal tasks, and long-running task execution with multi-step agent workflows.

Key Performance Improvements

GPT-5.4 demonstrates significant performance gains across various benchmarks. In the GDPval test, which assesses practical skills across 44 occupations, GPT-5.4 Thinking outperformed humans in 83.0% of cases, compared to 70.9% for GPT-5.2 Thinking and 38.8% for GPT-5.1 Thinking. Mashable details these improvements.

Category Test GPT-5.4 GPT-5.4 Pro GPT-5.3-Codex GPT-5.2 GPT-5.2 Pro
Professional GDPval 83,0 % 82,0 % 70,9 % 70,9 % 74,1 %
Finance FinanceAgent v1,1 56,0 % 61,5 % 54,0 % 59,5 %
Coding SWE-Bench Pro (Public) 57,7 % 56,8 % 55,6 %

Expanded Capabilities and Context Window

GPT-5.4 introduces a 1 million token context window (experimentally available in Codex), enabling analysis of entire codebases, extensive document collections, and prolonged agent interactions within a single request. GPT-5.4 is the first mainline model with built-in computer-use capabilities, allowing agents to interact directly with software for task completion, verification, and correction. The model is also trained to support compaction, preserving key context during extended agent trajectories.

Availability

GPT-5.4 Thinking is currently available to ChatGPT Plus, Team, and Pro subscribers, replacing GPT-5.2 Thinking, which will be phased out in three months. GPT-5.4 is rolling out gradually in ChatGPT and Codex. The new models are also accessible via the API, with GPT-5.4 being more expensive but more efficient than its predecessor.

Related Posts

Leave a Comment