OpenAI has announced the rollout of GPT-5.4, its latest artificial intelligence model, across ChatGPT, its developer API and Codex. According to the company, the model combines improvements in reasoning, coding and agent-driven workflows to handle complex professional tasks more effectively.
Within ChatGPT, the model will appear as GPT-5.4 Thinking, replacing GPT-5.2 Thinking for paid users. The company said the update aims to help the model complete real-world tasks more efficiently, including work involving spreadsheets, presentations and documents.
OpenAI also introduced GPT-5.4 Pro, a higher-performance version of the model designed for more demanding workloads and complex tasks.
GPT-5.4: What’s new
According to OpenAI, GPT-5.4 integrates the coding capabilities previously introduced with GPT-5.3 Codex while improving performance across reasoning, knowledge work and software environments. The company said the model is designed to reduce back-and-forth interactions by producing more accurate outputs in a single attempt.
On internal and third-party benchmarks cited by OpenAI, GPT-5.4 showed improvements over GPT-5.2 across several areas. On GDPval, a benchmark measuring the ability of AI systems to perform professional knowledge work, the model matched or exceeded industry professionals in 83 per cent of comparisons, compared with 70.9 per cent for GPT-5.2.
OpenAI also said the model performs better on tasks involving spreadsheets and presentations. In internal tests simulating work done by junior investment banking analysts, GPT-5.4 achieved an average score of 87.3 per cent in spreadsheet modelling tasks, compared with 68.4 per cent for GPT-5.2.
The company also claims the model is more reliable than previous versions. In tests involving prompts where users had previously flagged factual errors, GPT-5.4 generated responses that were 18 per cent less likely to contain mistakes compared with GPT-5.2.
GPT-5.4 Thinking: What’s new in ChatGPT
In ChatGPT, OpenAI introduces GPT-5.4 Thinking, which brings features aimed at making long or complex responses easier to guide.
The company said GPT-5.4 Thinking improves deep web research, particularly for highly specific queries that require information from multiple sources. OpenAI claims the model can maintain context over longer reasoning chains, helping it produce more coherent responses for complex questions.
Agent capabilities, tool search and computer use
OpenAI said GPT-5.4 introduces stronger support for agent-style workflows, where AI systems can complete tasks across multiple tools and software environments.
The model is the company’s first general-purpose system with native computer-use capabilities, allowing AI agents to interact with applications through screenshots, mouse movements and keyboard inputs. OpenAI said this enables agents to navigate websites, manage files and perform multi-step workflows across software systems.
Developers can access these capabilities through updated computer-use tools in the API.
OpenAI also introduced tool search, which allows the model to dynamically retrieve tool definitions when needed instead of loading all available tools into a prompt at once. According to the company, this approach can reduce token usage and improve efficiency when AI agents operate across large tool ecosystems.
In addition, agentic tool calling has been improved. OpenAI said GPT-5.4 can more accurately decide when to use external tools during reasoning tasks, which helps agents complete multi-step workflows with fewer interactions and lower latency.
GPT-5.4 Pro: Higher-performance version
Alongside the base model, OpenAI introduced GPT-5.4 Pro, which the company describes as a higher-performance version intended for complex tasks requiring maximum capability.
On the BrowseComp benchmark, which evaluates how well AI agents search and synthesise information from the web, GPT-5.4 Pro achieved an accuracy score of 89.3 per cent, compared with 82.7 per cent for the standard GPT-5.4 model.
The company said GPT-5.4 Pro is available to ChatGPT Pro and Enterprise users, as well as developers through the API.
Availability
GPT-5.4 has been rolled out across ChatGPT, the API and Codex. In ChatGPT, the model is available to Plus, Team and Pro users as GPT-5.4 Thinking, replacing GPT-5.2 Thinking.
Developers can access the model through the API under the name gpt-5.4, while gpt-5.4-pro is available for higher-performance workloads.
OpenAI said GPT-5.2 Thinking will remain accessible in ChatGPT under the Legacy Models section for three months before being retired.
GPT-5.3 Instant
In related news, OpenAI recently introduced
GPT-5.3 Instant, an update to the default model powering ChatGPT, aimed at improving everyday conversations. The company said the update focuses on making responses more fluid, direct and relevant, while reducing overly cautious language and unnecessary refusals. According to OpenAI, earlier models sometimes declined questions that could have been answered safely or used long disclaimers before responding. With GPT-5.3 Instant, the model is designed to provide clearer answers with fewer defensive preambles and a more natural conversational tone.
OpenAI also claims improvements in factual reliability and writing quality. The company said GPT-5.3 Instant reduces hallucinations across domains such as medicine, law and finance, with internal tests showing up to a 26.8 per cent drop in hallucination rates when web access is used. The model also aims to produce more coherent writing and maintain a consistent conversational style. GPT-5.3 Instant is now available to all ChatGPT users and developers through the API as gpt-5.3-chat-latest, while GPT-5.2 Instant will remain accessible for paid users in the Legacy Models section until June 3, 2026.
 "OpenAI releases GPT-5.4 Thinking and Pro in ChatGPT: How is it different")
