OpenAI launches GPT-5.4: a more capable AI for professional work | Keryc
OpenAI introduces GPT-5.4, a jump designed for professional work: more accurate, more efficient, and with new skills to operate applications and handle long documents. Can you imagine an assistant that plans its own reasoning while you work with it? That's part of what's new.
What GPT-5.4 brings
GPT-5.4 combines improvements in reasoning, coding and tool use into a single model. It's available as GPT-5.4 Thinking in ChatGPT, as gpt-5.4 in the API, and also in Codex. There's also a Pro version (gpt-5.4-pro) for extremely complex tasks that demand top performance.
In plain terms: the model gives more accurate results with fewer back-and-forths, and can handle real workflows that previously needed a lot of human oversight.
Key new features that matter for professionals
Visible planning: in ChatGPT, GPT-5.4 Thinking can show an initial plan of its reasoning so you can adjust it while it responds. Want to change the route halfway through? Now you can.
Better handling of documents and sheets: OpenAI emphasizes spreadsheets, presentations and documents. In tasks like a junior investment banking analyst, GPT-5.4 rises to 87.3% accuracy versus 68.4% for GPT-5.2. Generated presentations are preferred by human evaluators for aesthetics and visual variety.
Native computer control: for the first time the model can operate applications, interact with screens (clicks, keys) and use libraries like Playwright to run complex flows. That enables agents that not only think, but do.
Huge context window: it supports up to 1 million tokens in experimental uses, which helps plan and execute tasks that need to remember lots of context over time.
Tool search: the API introduces tool search, which avoids loading all tool definitions on every request, saving tokens and speeding up responses when many connectors are available.
Performance and concrete examples
GPT-5.4 improves on real tests and relevant benchmarks:
GDPval (professional work): 83.0% vs 70.9% for GPT-5.2.
BrowseComp (persistent web search): 82.7% with the Thinking version; the Pro variant reaches 89.3%.
OSWorld-Verified (desktop use with screenshots): 75.0%, higher than GPT-5.2 and above human performance on that evaluation.
What does that mean for you? Fewer manual corrections, more complete answers for tasks that require gathering information from many sources, and agents able to log into portals, fill forms and verify results almost like a human.
Speed, efficiency and price
OpenAI says GPT-5.4 is their most efficient model in terms of reasoning tokens: it solves problems using significantly fewer tokens than GPT-5.2, which can lower costs and speed up responses. Still, the price per token is higher than GPT-5.2 to reflect the improvements.
Indicative pricing (summary):
gpt-5.2: lower price per token.
gpt-5.4: higher price per token, but greater efficiency on real tasks.
Pro versions: higher fees for critical loads and prioritized latency.
In practice, paying more per token can pay off if the model needs fewer tokens to finish the job.
Security and controls
OpenAI keeps a cautious stance: GPT-5.4 is classified as high cyber capability under their readiness framework. In environments with zero data retention (ZDR) additional mechanisms apply, like asynchronous locking and strong access controls. They also published studies on monitoring reasoning (Chain-of-Thought) and believe GPT-5.4 does not hide its thought process, which helps detect potential misuse.
Practical tools available starting today
In ChatGPT: GPT-5.4 Thinking comes to Plus, Team and Pro; GPT-5.4 Pro for Pro and Enterprise. GPT-5.2 Thinking will remain as legacy for three months.
In the API and Codex: gpt-5.4 is already available; the API includes renewed tools for computer use and updated documentation.
Add-ins and demos: there's an Excel add-in and an experimental skill called Playwright (Interactive) to debug visual apps while you build them.
What does this mean for your work or project?
If you work with complex spreadsheets, corporate presentations or need to automate processes that involve multiple web apps, GPT-5.4 promises to reduce friction. For agent and automation developers, the improvements in tool control and native computer use are especially relevant.
Does this mean you no longer need to review anything? No. Although error rates drop and factuality improves, it's still wise to validate critical results, especially on legal, financial or security matters.
GPT-5.4 is a practical step forward: more capable, faster and designed to fit into real workflows. It's a tool that cuts repetitive work and speeds up complex deliverables, while still requiring human oversight on sensitive tasks.
