OpenAI released GPT-5.4. This is the latest frontier model in the GPT-5 series, described by OpenAI as their “most capable and efficient” version yet, specifically optimized for professional work, knowledge tasks, coding, and agentic workflows.
It unifies advances from previous releases like GPT-5.3-Codex for coding into a single system. For the first time in a general-purpose OpenAI model, it can control a computer like a human — navigating interfaces, clicking, typing, browsing, and working across apps. This enables true AI agents for automation.
1 million token context window (in API/Codex): Supports very long contexts (up to ~922K input / 128K output in some configs), great for complex, multi-step tasks without losing track. Improved efficiency and factuality: Uses significantly fewer tokens (up to 47% less on some tasks), runs faster, hallucinates less (33% fewer false claims vs. GPT-5.2, 18% fewer overall errors), and retains context better during long “thinking” sessions.
In ChatGPT, you can now interrupt the model while it’s thinking and generating and adjust instructions or direction on the fly (rolling out on web/Android now, iOS soon). Enhanced reasoning, coding, and tools: State-of-the-art on benchmarks for professional tasks like spreadsheets, documents, presentations, web research.
Register for Tekedia Mini-MBA edition 20 (June 8 – Sept 5, 2026).
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Register for Tekedia AI Lab.
It supports “reasoning.effort” levels and excels at agentic planning, execution and verification. GPT-5.4 Pro — Max performance for the hardest tasks. GPT-5.4 Thinking is rolling out to Plus, Team, and Pro users (Enterprise/Edu via admin settings). Pro version for higher tiers. API: Available immediately as gpt-5.4 and gpt-5.4-pro.
Codex: Integrated for coding and agent use. Legacy models like GPT-5.2 Thinking stay accessible for ~3 months until June 2026. Pricing NotesIt’s positioned as premium: higher per-token costs than predecessors; input and output rates reflect frontier status, with multipliers for very long contexts >272K tokens).
But efficiency gains (fewer tokens needed) can offset this for many workflows. This comes just days after GPT-5.3 Instant showing OpenAI’s rapid iteration pace. Early user and dev feedback highlights big jumps in practical agent and autonomy use cases, though some note it’s still evolving amid broader company context.
Native computer use in GPT-5.4 refers to OpenAI’s built-in, state-of-the-art capability that allows the model to directly interact with and control a computer interface — much like a human user would. This is a major advancement toward truly autonomous AI agents, and it’s the first time OpenAI has integrated this natively into a general-purpose frontier model.
The model operates in a visual + action loop:It receives screenshots or screen captures of the current interface. It analyzes what’s on screen using its vision understanding. It decides on the next action and outputs structured commands, such as: Moving and clicking the mouse at specific coordinates.
Typing text or keystrokes. Scrolling, dragging elements, or navigating menus. Your code or harness (the surrounding software) executes those actions in the real environment. It gets the updated screenshot back and repeats — forming a closed loop of observe ? plan ? act ? verify ? correct.
This enables multi-step, real-world workflows without needing pre-built APIs for every tool. GPT-5.4 excels at both: Code-based control — Writing automation scripts using Playwright for browsers Direct low-level control — Issuing raw mouse and keyboard events based purely on visual input.
Developers can steer its behavior through prompts, set custom safety rules; requiring user confirmation for risky actions like deleting files or making payments, and adjust risk tolerance. This beats GPT-5.2’s 47.3% and even surpasses average human performance (72.4%).
It also leads on related agentic benchmarks like WebArena and BrowseComp, showing big gains in reliability for long-horizon tasks. Real-World ExamplesAutomate filling out forms across multiple websites/apps. Pull data from Excel/Google Sheets ? analyze it ? generate a report/presentation. Navigate file systems, open documents, edit them, and save changes.
Debug software by controlling an IDE, running tests, and fixing issues in a loop. Handle repetitive office workflows; data entry, invoice processing, research + summarization across tools. In practice, this turns GPT-5.4 into something closer to a digital employee that can “use” your computer directly, rather than just suggesting steps for you to follow.
It’s designed for professional/enterprise use, with configurable safeguards to prevent misuse. This feature marks a big step in the shift from chat-based AI to action-taking agents — enabling more autonomous, end-to-end automation in knowledge work.