New Gpt-5.4 Clobbers Humans On Pro-level Work In Openai's Tests - By 83%

OpenAI

Follow ZDNET: Add america arsenic a preferred source connected Google.

ZDNET's cardinal takeaways

• GPT-5.4's 83% people suggests AI rivals master professionals.
• Tests span 9 industries and 44 real-world occupations.
• New capabilities boost coding, tools, and machine control.

It seems for illustration only yesterday that OpenAI released its GPT-5.2 model to nan world. In fact, it's been little than 3 months. Thursday, OpenAI is releasing nan reasoning exemplary of GPT-5.4.

Also: How to move from ChatGPT to Claude: Transferring your memories and settings is easy

What precisely does that mean? In this article, I'll concisely touch connected nan charismatic announcement and readiness details, and past I'll dive into what I deliberation is nan astir startling detail: GPT-5.4 tin lucifer aliases outperform quality professionals 83% of nan time, according to OpenAI.

(Disclosure: Ziff Davis, ZDNET's genitor company, revenge an April 2025 suit against OpenAI, alleging it infringed Ziff Davis copyrights successful training and operating its AI systems.)

Availability specifications

OpenAI says GPT-5.4 is "the astir tin and businesslike frontier exemplary for analyzable master work." Within ChatGPT, nan institution calls this exemplary GPT 5.4 Thinking. There are besides releases for nan API, wrong nan Codex programming tool, and successful a GPT-5.4 Pro version.

Also: 10 ChatGPT Codex secrets I only learned aft 60 hours of brace programming pinch it

In position of wide performance, nan institution says that GPT-5.4 is "18% little apt to incorporate errors, and individual claims are 33% little apt to beryllium mendacious compared to GPT-5.2, based connected prompts wherever users antecedently flagged actual mistakes."

It's ever bully erstwhile an highly powerful artificial intelligence makes worldly up little frequently.

As for availability, nan institution will connection GPT-5.4 via API connected Friday. It will beryllium "rolling out" crossed ChatGPT paid tiers and successful Codex, which presumably intends it will show up reasonably soon for astir users.

But what astir GPT-5.3?

It gives maine nary joyousness to opportunity this, but OpenAI's naming conventions springiness maine a headache. When it comes to naming, it feels for illustration it fired each its knowledgeable merchandise managers and replaced them pinch a GPT-3.5 lawsuit from 2022.

So, OK, OpenAI released GPT-5.3-Codex past month. That's nan first type of Codex that utilized itself to thief build itself. Skynet, anyone?

Then, 2 days ago...two days ago it released GPT-5.3 Instant. This, according to nan company, "makes mundane conversations much consistently adjuvant and fluid." It's disposable to each users of ChatGPT. In nan API, it's released arsenic gpt-5.3-chat-latest. Not gpt-5.3-chat-instant, because that would make excessively overmuch sense.

And now, we person GPT-5.4. So successful nan abstraction betwixt Tuesday and Thursday, OpenAI has released a GPT-5.3 and a GPT-5.4 model. You'd person to beryllium an AI to support way of it all.

Because specified crimes against coherent versioning make maine twitchy, I had to inquire nan OpenAI communications squad astir it. They were diligent and benignant capable to answer:

GPT-5.4 is our first mainline reasoning exemplary that incorporates nan frontier coding capabilities of gpt-5.3-codex, and that is rolling retired crossed ChatGPT, nan API, and Codex. We're calling it GPT-5.4 to bespeak that jump, and to simplify nan prime betwixt models erstwhile utilizing Codex. Over time, you tin expect our Instant models and Thinking models to germinate astatine different speeds.

I still don't for illustration it. If Instant and Thinking are really 2 abstracted products, they should person wholly abstracted versioning. 5.3 and 5.4 are excessively adjacent and excessively confusing. If they're considered to beryllium different variants of nan aforesaid product, they should stock type numbers.

Also: Is ChatGPT Plus still worthy your $20? I compared it to nan Free, Go, and Pro plans

But hey. OpenAI is worthy something connected nan bid of $840 billion, and I ain a 14-year-old Ford. What do I know? Let's move connected to nan portion wherever we each interest astir our occupation security.

Testing real-world AI expertise

In September, OpenAI introduced a new AI information trial called GPTval. It's a trial designed to measurement really good AI models execute doing "economically valuable, real-world tasks."

The trial measures capacity successful 9 industries and 44 occupations. OpenAI chose nan industries based connected those contributing 5% aliases much to nan US gross home product. Each manufacture has unsocial occupations. For nan test, nan institution selected up to 5 occupations, choosing those that had little than 40% beingness aliases manual work, and which dress up those jobs pinch nan highest full wages and astir wide compensation.

Also: I stopped utilizing ChatGPT for everything: These AI models hit it astatine research, coding, and more

It fundamentally picked a cross-section of knowledge-related jobs wherever AI could person nan astir effect "on real-world productivity." The intent was that nan GPT models could thief professionals get much done, but it's not excessively large a leap to infer that these occupations are besides nan astir astatine consequence from AI replacement aliases augmentation.

Here's really those occupations fresh into their industries.

• Finance and insurance: Customer work representatives, financial and finance analysts, financial managers, individual financial advisors, securities, commodities, and financial services income agents
• Retail trade: Pharmacists, first-line supervisors of unit income workers, wide and operations managers, backstage detectives and investigators
• Wholesale trade: Sales managers, bid clerks, first-line supervisors of non-retail income workers, income representatives (wholesale and manufacturing, isolated from method and technological products), income representatives (wholesale and manufacturing, method and technological products)
• Real property and rental and leasing: Concierges, property, existent estate, and organization relation managers, existent property income agents, existent property brokers, antagonistic and rental clerks
• Government: Recreation workers, compliance officers, first-line supervisors of constabulary and detectives, administrative services managers, child, family, and schoolhouse societal workers
• Manufacturing: Mechanical engineers, business engineers, buyers and purchasing agents, shipping, receiving, and inventory clerks, first-line supervisors of accumulation and operating workers
• Professional, scientific, and method services: Software developers, lawyers, accountants and auditors, machine and accusation systems managers, task guidance specialists
• Health attraction and societal assistance: Registered nurses, caregiver practitioners, aesculapian and wellness services managers, first-line supervisors of agency and administrative support workers, aesculapian secretaries and administrative assistants
• Information: Audio and video technicians, producers and directors, news analysts, reporters, and journalists, movie and video editors, editors

I could get picky astir which occupations are nan astir impactful successful nan various industries, but this action is simply a bully 1 for testing exemplary capacity overall.

The tests themselves are absorbing successful some really they are constructed and really they are measured.

OpenAI worked pinch knowledgeable professionals successful each business to create a group of tasks that "reflect their day-to-day work." The task sets each went done galore rounds of master reappraisal and resulted successful a bid of afloat reviewed, analyzable tasks per industry.

One of nan manufacturing technologist tasks, for example, involves nan creation of a jig (guides a tool) aliases a fixture (holds nan work) to simplify nan reeling successful and reeling retired of a cablegram spool for underground mining operations.

Also: This elemental ChatGPT instrumentality helps you spot scams earlier you click aliases respond

Grading for each of these tests was done by quality professionals successful each of nan occupations. The graders weren't told whether nan results were from nan AI, aliases from different professionals successful their fields.

Additionally, OpenAI built an automated grading strategy based connected nan activity of nan quality graders, truthful that nan humans don't person to return their clip grading each loop of nan AI model. I'm judge OpenAI constructed this automated strategy pinch each due safeguards, but I interest that immoderate level of inherent bias mightiness beryllium imaginable erstwhile letting an AI people nan capacity of an AI.

Ethan Mollick, subordinate professor and co-director of nan Generative AI Lab astatine Wharton, describes nan GDPval test arsenic "probably nan astir economically applicable measurement of AI ability."

83% of nan clip

The velocity of betterment is insane. GPT-5.1 was released successful November and had a GDPval people of 38.8%. In December, conscionable a period later, GPT-5.2 capacity exploded to astir double that, to 70.9%.

Professor Mollick described nan value of GDPval moving connected GPT-5.2. He said, "In head-to-head title pinch quality experts connected tasks that require 4-8 hours for a quality to do, GPT-5.2 wins 71% of nan clip arsenic judged by different humans."

Now, successful early March, little than 3 months aft GPT-5.2, GPT-5.4 matches aliases exceeds nan capacity of quality professionals 83% of nan time!

Also: How to study ChatGPT successful an hr - for free

In different words, almost each clip nan aforesaid task was fixed to an knowledgeable quality pro and GPT-5.4, nan AI either kept up pinch aliases blew past nan knowledgeable quality pro, astatine least, according to its grader, which whitethorn person been quality aliases AI.

Sit pinch that for a fewer minutes. We're not conscionable talking astir programming tasks. We're talking astir a wide scope of industries and a wider scope of high-value occupations.

According to Daniel Swiecki, caput of Artificial Intelligence Solutions astatine Walleye Capital, "On our toughest soul finance and Excel evaluations, GPT-5.4 outperformed anterior models, improving accuracy by 30 percent points. This measurement alteration successful reliability materially expands our automation of exemplary updates and script analyses for basal investors."

The freaky point is this benignant of capacity could return america successful 2 directions. On nan 1 hand, it could thief augment quality pros, giving knowledgeable folks nan expertise to get much done, faster. On nan different hand, it could good beryllium seen arsenic nan harbinger of a clip erstwhile nan AI is simply replacing nan humans successful high-value, high-skill jobs.

The early is astir apt not going to beryllium each 1 aliases each nan other. But moreover arsenic OpenAI takes a triumph thigh for its latest release, those of america who support our families based connected a life of accomplishment building wrong those professions person to stone backmost connected our heels, return deep, worried breaths, and dream for nan best.

Speaking personally, my attack has been to study each I can, arsenic quickly arsenic I can, and usage AI arsenic overmuch arsenic I can. That helps maine picture each of this to you, but it besides helps maine augment my individual productivity utilizing AI resources, peculiarly for programming.

Also: I'm a ChatGPT powerfulness user: Here are 7 useful settings that are turned disconnected by default

But I worry. AI slop is simply a existent thing, and arsenic AI slop increases much and much successful quality, each of america will beryllium competing pinch a elephantine superbrain that ne'er sleeps, ne'er eats, and is improving astatine almost supernatural speed.

More capabilities

In summation to wide performance, GPT-5.4 improves connected different halfway capabilities.

• Tool use: GPT-5.4 improves really AI agents prime and usage outer tools, enabling them to complete multi-step workflows much accurately and efficiently while reducing token usage.
• Computer vision: The caller exemplary enhances ocular understanding, allowing it to amended construe analyzable images, parse documents, and logic astir ocular accusation pinch higher accuracy.
• Computer usage capabilities: Within nan API and Codex, GPT-5.4 introduces autochthonal computer-use abilities that fto agents interact pinch package systems done screenshots, keyboard and rodent commands, and automated workflows crossed applications.
• Coding: GPT-5.4 combines nan coding strengths of GPT-5.3-Codex pinch improved reasoning and instrumentality use, helping developers build, debug, and iterate connected analyzable package tasks much effectively.

Stay tuned. GPT-5.4 Thinking will beryllium successful your ChatGPT interface shortly. Let nan title begin.

What do you think?

What do you deliberation astir GPT-5.4's declare that it tin lucifer aliases outperform quality professionals 83% of nan time? Does that look for illustration a meaningful benchmark for real-world work?

Also: The champion AI chatbots of 2026: Expert tested and reviewed

Have you started integrating AI into your ain master workflow? If so, wherever does it thief nan astir aliases autumn short? Looking ahead, do you spot devices for illustration this mostly augmenting quality expertise, aliases yet replacing parts of it?

Share your thoughts and experiences successful nan comments below.

You tin travel my day-to-day task updates connected societal media. Be judge to subscribe to my play update newsletter, and travel maine connected Twitter/X astatine @DavidGewirtz, connected Facebook astatine Facebook.com/DavidGewirtz, connected Instagram astatine Instagram.com/DavidGewirtz, connected Bluesky astatine @DavidGewirtz.com, and connected YouTube astatine YouTube.com/DavidGewirtzTV.