We Tried GPT-5.4 And it is Not Your Regular AI Chatbot Anymore

OpenAI is out with a significant replace, constructing on its GPT-5 collection with the all-new GPT-5.4. Launched as GPT-5.4 Considering, the mannequin will even include a GPT-5.4 Professional model for these searching for “most efficiency” on sophisticated duties. Even the bottom model comes with a plethora of enhancements over the outgoing GPT-5.2. These upgrades vary throughout reasoning, coding, and agentic workflows, together with some nifty little options that customers are positive to like.

As an illustration, OpenAI says that the GPT-5.4 Considering will allow you to alter the course of its considering in the midst of a response. This implies extra applicable outcomes on your queries. Aside from this, it carries enhancements in deep internet search and bigger context home windows. All in all, higher-quality, extra correct solutions in much less time.

Right here, we discover all these options and benchmark performances of the brand new GPT-5.4 Considering intimately, beginning proper with what the AI mannequin is all about.

Additionally learn: New Replace Makes GPT-5.3 Immediate Extra Helpful For On a regular basis Duties

What’s GPT-5.4?

Identical to with each new AI mannequin, OpenAI makes use of the time period “most succesful and environment friendly frontier mannequin” whereas introducing GPT-5.4 in its weblog. Nevertheless, there’s a follow-up time period that sheds a a lot brighter mild on its nature. The adjectives above are used particularly in reference to “skilled work.” Which signifies that the GPT-5.4, not like earlier fashions that primarily pushed conversational intelligence, comes as a devoted AI mannequin for professionals.

For this, it brings enhancements in reasoning, coding, and agentic workflows right into a single system that’s meant to deal with actual duties throughout software program instruments and digital environments. Yet one more spotlight is its assist for a large context window of as much as 1 million tokens. This permits the mannequin to course of lengthy paperwork, datasets, and multi-step workflows with out shedding observe of the duty. On high of that, OpenAI says GPT-5.4 is its most token-efficient reasoning mannequin but, utilizing considerably fewer tokens than GPT-5.2 to reach at solutions.

The listing of options doesn’t finish right here. Subsequent, let’s take a look at all the important thing options that the GPT-5.4 carries.

Additionally learn: I Tried GPT 5.2 and That is How It Went..

Key Highlights of the GPT-5.4 Household

Listed here are the important thing highlights of the GPT-5.4 household.

1. Native Pc Use and Stronger Imaginative and prescient Capabilities

One of many greatest upgrades in GPT-5.4 is its capability to work together with computer systems and visible interfaces extra successfully. The mannequin introduces native computer-use capabilities, permitting AI brokers to function software program environments and execute workflows throughout functions. Mix this with stronger imaginative and prescient skills, and the GPT-5.4 can higher interpret screenshots, paperwork, and UI components. This permits it to navigate methods, extract info, and full duties that require each visible understanding and motion throughout instruments.

2. Smarter Instrument Discovery and Utilization

GPT-5.4 additionally improves how fashions work together with massive ecosystems of instruments and connectors. It introduces one thing known as “instrument search,” which helps the mannequin establish and use the proper instruments inside complicated environments. As an alternative of relying solely on predefined integrations, GPT-5.4 can dynamically uncover the instruments wanted to finish a process. This makes it simpler to construct AI methods that work throughout a number of companies with out sacrificing reasoning functionality.

3. Improved Efficiency on Information Work

A serious focus of GPT-5.4 is dealing with skilled information work extra reliably. The mannequin exhibits stronger efficiency on duties involving spreadsheets, displays, and lengthy paperwork, the place sustaining context and accuracy is vital. In response to OpenAI’s evaluations, GPT-5.4 considerably improves output high quality on these kinds of duties, producing extra polished outcomes whereas requiring fewer corrective prompts from customers.

4. Stronger Coding and Developer Workflows

GPT-5.4 additionally builds on the coding strengths launched in GPT-5.3-Codex. The mannequin maintains sturdy efficiency on software program engineering benchmarks whereas bettering its capability to deal with longer improvement workflows. This permits it to help with debugging, writing code throughout a number of information, and coordinating duties that require reasoning throughout massive codebases.

5. Better Management By means of Steerability

One other enchancment comes within the type of higher steerability, particularly in ChatGPT. With GPT-5.4 Considering, the mannequin can current an upfront reasoning plan earlier than producing its remaining output. This permits customers to information the path of the response whereas it’s nonetheless being generated, lowering the necessity for repeated prompts and making complicated duties simpler to handle.

6. Expanded Cyber Security Stack

Lastly, GPT-5.4 introduces an expanded cyber security stack designed to cut back dangerous or unsafe outputs. OpenAI has strengthened safeguards in opposition to malicious use whereas bettering the mannequin’s capability to refuse inappropriate requests. These upgrades intention to make the system extra dependable and safe when deployed throughout enterprise and developer environments.

With these claims, OpenAI has additionally shared some sturdy benchmark efficiency outcomes of the GPT-5.4. Let’s take a look at them right here.

GPT-5.4 – Benchmark Efficiency

Benchmarks are sometimes the place the actual story of a brand new AI mannequin begins to indicate up. And within the case of GPT-5.4, the numbers recommend that OpenAI’s deal with skilled work isn’t just advertising language. Throughout a number of classes, from finance and coding to instrument utilization and reasoning, the mannequin persistently edges previous its predecessors.

Take skilled information duties, for instance. On the GDPval benchmark, GPT-5.4 scores 83%, a noticeable leap from 70.9% for GPT-5.2. An identical pattern seems in monetary modelling duties, the place GPT-5.4 achieves 87.3% accuracy, in comparison with 68.4% for GPT-5.2. These benchmarks simulate real-world skilled work similar to analysing spreadsheets, constructing monetary fashions, and answering office-related queries. In less complicated phrases, the mannequin appears much better outfitted to deal with the sorts of duties professionals truly cope with every day.

The enhancements should not restricted to workplace work. In laptop use and imaginative and prescient duties, GPT-5.4 data a 75% rating on the OSWorld-Verified benchmark, dramatically larger than GPT-5.2’s 47.3%. This means the mannequin is considerably higher at interacting with laptop interfaces, understanding visible inputs, and finishing workflows throughout functions. On tool-use benchmarks like BrowseComp, GPT-5.4 reaches 82.7%, indicating stronger efficiency when the mannequin has to seek out, choose, and use the proper instruments to finish a process.

Even in conventional coding and reasoning benchmarks, the positive factors are regular. GPT-5.4 barely improves the SWE-Bench Professional rating to 57.7%, constructing on the already sturdy coding capabilities launched in GPT-5.3-Codex. In the meantime, on summary reasoning exams like ARC-AGI, GPT-5.4 jumps to 93.7%, far forward of GPT-5.2’s 86.2%. Put collectively, these numbers reinforce the very purpose of GPT-5.4’s being: an AI mannequin designed not simply to talk, however to assume by means of complicated issues and full actual work throughout domains.

Now that we all know how succesful GPT-5.4 is, how will we entry it?

GPT-5.4: Availability and Pricing

The excellent news is that the GPT-5.4 is already rolling out throughout ChatGPT, the API, and Codex. The not-so-good information for some – it is going to be restricted to the Plus, Workforce, and Professional customers for now. OpenAI says that the brand new mannequin will seem as GPT-5.4 Considering underneath the mannequin picker on ChatGPT.

For builders, GPT-5.4 is already dwell within the API as gpt-5.4, whereas the higher-performance gpt-5.4-pro variant is offered for workloads that require most reasoning energy. In the meantime, Enterprise and Edu customers can allow early entry by means of admin settings, and Codex customers will see GPT-5.4 built-in into their improvement workflows as effectively. Here’s a take a look at its API pricing:

With the arrival of the GPT-5.4, OpenAI can be gearing as much as bid farewell to the older one, i.e. GPT-5.2. GPT-5.2 Considering will stay accessible for paid customers underneath the Legacy Fashions part for the following three months, after which it is going to be retired on June 5, 2026.

Now you already know the place to get it, here’s a glimpse of it in real-world motion.

Additionally learn: The best way to Use ChatGPT? A Easy Information for Inexperienced persons

GPT-5.4 Considering: Palms-on

Since GPT-5.4 is positioned for skilled work, I focused the three areas in my hands-on the place it claims the largest enhancements. These are:

Information work (paperwork, evaluation, structured considering)

Coding and technical workflows

Agentic workflows / tool-based duties

Take a look at the outputs for every and expertise what GPT-5.4 brings to the desk.

1. Information work

Immediate:

I’m sharing a report. Your process is to:

– Summarize the doc in underneath 200 phrases.
– Extract the 5 most essential insights.
– Determine any assumptions or weak arguments within the textual content.
– Recommend two actionable suggestions primarily based on the evaluation.
– Construction your reply clearly underneath headings.

Output:

  

As we are able to see, GPT-5.4 dealt with the long-document fairly effectively. The abstract was concise and displays the core argument of the paper. I discover no pointless particulars in anyway – an enormous plus. The important thing insights have been logically extracted and mirrored the doc’s central themes. Within the assumptions part, the mannequin confirmed good vital considering, mentioning sensible considerations round battery progress, prices, and public acceptance. Lastly, the suggestions – suggesting pilot programmes and ecosystem improvement moderately than unrealistic large-scale deployment – appear sensible and straight derived from the evaluation.

2. Coding and developer workflow

Immediate:

I need to construct a Python script that does the next:

Scrapes the newest AI information headlines from 3 know-how web sites.

Cleans and deduplicates the headlines.

Makes use of a easy sentiment classifier to label every headline as optimistic, impartial, or unfavorable.

Shops the ends in a CSV file.

First define the structure of the script. Then write the whole Python code with feedback.

Output:

  

The response performs effectively on workflow planning as a result of it lays out the pipeline in a logical order: supply setup, scraping, cleansing, sentiment tagging, and CSV export. This makes the end-to-end circulation simple to observe. On code execution, it’s sturdy for a beginner-to-intermediate use case, with runnable code, error dealing with, modular features, and feedback, although the scraping layer nonetheless will depend on fragile CSS selectors and a really primary sentiment technique.

I discover the structure cleanliness one of many higher components of the reply. Obligations are separated neatly, features are modular, and the script is simple to increase, even when it stops wanting a extra production-grade design with config information, logging, and reusable scraper abstractions. All in all, the output demonstrates a powerful use-case of GPT-5.4 for coding and developer workflows.

3. Agentic workflow

Immediate:

You might be advising a startup deciding between three AI enterprise concepts:

An AI-powered monetary analysis assistant

An AI doc automation platform for legislation companies

An AI agent that automates e mail workflows

Consider these concepts primarily based on:

market dimension

problem of execution

aggressive panorama

monetization potential

Present a structured comparability desk and suggest the best choice.

Output:

  

I personally just like the structured reasoning right here, as a result of it compares all three concepts by means of the identical 4 enterprise lenses after which converts that comparability into a transparent suggestion. The logic thus turns into simple to observe. Its readability of thought is robust too: the distinctions between “broad however crowded,” “excessive worth however laborious,” and “centered with sturdy ROI” are communicated cleanly with out pointless jargon.

The standard of research is stable general, particularly in highlighting execution problem, purchaser willingness to pay, and aggressive stress. What I did notice is that it stays at a strategic degree and will have been even stronger with sharper startup-stage nuances like go-to-market velocity, founder-market match, or preliminary wedge technique. General, this can be a superb instance of GPT-5.4 producing a enterprise reply that feels organised, commercially conscious, and instantly helpful moderately than simply generically clever.

Conclusion

We’re means past AI chatbots now. With GPT-5.4, OpenAI is clearly focusing on a extremely reliable co-worker for all kinds {of professional} duties. And the capabilities of the AI mannequin, as we have now seen with our exams, are excellent in these regards.

From long-document evaluation to agentic workflows, even in my restricted use of GPT-5.4 until now, it looks like a mannequin constructed for individuals who need AI to really assist them get work performed. It might not change every little thing in a single day, but it surely does push AI one step nearer to what we truly need out of it, to assist us with real-world duties, and never simply our questions.