News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnGPT-5.4

GPT-5.4

OpenAI's March 2026 large language model with computer use and deep research.

GPT-5.4 is a large language model that OpenAI put out on 5 March 2026, following GPT-5.2. It arrived first as Thinking and Pro editions, with mini and nano versions added later in the month. The company reports fewer factual mistakes than its predecessor and stronger results on professional tasks, plus built-in control of desktop software.

Share on WhatsApp

  • Released 5 March 2026 as Thinking and Pro; mini and nano followed on 17 March
  • OpenAI reports 33% fewer factual errors than GPT-5.2
  • OSWorld-Verified score of 75%, against 47.3% for GPT-5.2 and 72.4% for a human
  • Can operate desktop software and does deeper research work
  • Mini and nano cost four times their GPT-5 equivalents on the API

What it is

GPT-5.4 stands for Generative Pre-trained Transformer 5.4. It is a large language model, meaning a system trained on text that answers in text. OpenAI released it on 5 March 2026 as the successor to GPT-5.2. The launch covered two editions, Thinking and Pro, and both sit behind a paid subscription. A further pair, mini and nano, arrived on 17 March. OpenAI aimed the improvements at professional work: fewer invented facts, better handling of deep research, and the ability to operate a desktop computer on the user's behalf.

How it came about

OpenAI shipped GPT-5.4 Thinking and GPT-5.4 Pro together on 5 March 2026. Neither edition was opened to people on the free tier. Twelve days later the company added GPT-5.4 mini and GPT-5.4 nano. Mini became the one version free users could reach, while nano stayed behind the OpenAI API. The line sits between GPT-5.2 before it and GPT-5.5 after it.

What it can do

OpenAI's own testing found a 33% drop in made-up facts when the model is set against GPT-5.2. The gains were aimed at professional workflows. Two features stand out: the model can drive desktop software by itself, and its deep research work was strengthened. On OSWorld-Verified, a test of how well a language model operates a desktop environment, GPT-5.4 reached 75%. The same test gave GPT-5.2 47.3%, while a typical human scored 72.4%.

How to use it

Thinking and Pro are for paying subscribers only. Mini is the single edition free-tier users can try. Nano is restricted to the OpenAI API. On that API, both mini and nano cost four times as much as the matching GPT-5 versions. The weights are proprietary, so there is nothing to download.

Reception and debate

ZDNET liked how rarely GPT-5.4 Thinking invents information, but found that it sometimes strays from what the user asked. Vice singled out the stronger deep research and the fact that Thinking answers with fewer tokens than GPT-5.2 Thinking. The Decoder took issue with the pricing of mini and nano, which cost four times their GPT-5 counterparts on the API.

Questions people ask about GPT-5.4

What exactly is GPT-5.4?

It is a large language model from OpenAI, released on 5 March 2026. It writes text in response to text, and can also drive desktop software and carry out deep research. It sits after GPT-5.2 and before GPT-5.5 in the company's line.

Which versions exist and who can use them?

Four editions are listed: Thinking and Pro, both paid-only, then mini and nano from 17 March. Mini is the only one open to free-tier users. Nano can be reached solely through the OpenAI API.

How does it differ from GPT-5.2?

OpenAI measured a 33% cut in factual errors against GPT-5.2, with the work focused on professional tasks. On OSWorld-Verified, a desktop-operation test, it scored 75% versus 47.3% for GPT-5.2. Deep research was also improved.

Is it free?

Only mini is offered to free-tier users. Thinking and Pro need a paid subscription, and nano is limited to the API. On that API, mini and nano are four times the price of the equivalent GPT-5 models.

What have reviewers said?

ZDNET praised Thinking for rarely inventing facts but said it sometimes ignores instructions. Vice welcomed the deeper research and leaner token use. The Decoder objected to the API pricing of mini and nano.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "GPT-5.4", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.