OpenAI launches GPT-5.4 with a focus on professional work and autonomous agents

OpenAI launched this Thursday (5) GPT-5.4, its most capable and efficient frontier model aimed at professional work. Disponível in ChatGPT as GPT-5.4 Thinking, in the API and in Codex, the new system integrates recent advances in reasoning, coding and agentic workflows. Version Pro meets demands for maximum performance in complex tasks. The model performs professional activities with greater precision and less need for adjustments, benefiting users in corporate and development environments.
GPT-5.4 combines coding features from GPT-5.3-Codex with improvements to tools, software, and tasks involving spreadsheets, presentations, and documents. Ele delivers results aligned to requests with fewer repeated interactions. In ChatGPT, GPT-5.4 Thinking displays a pre-thought plan, allowing real-time adjustments for more accurate responses. Web search has gained depth in specific queries and maintains context in extended dialogues.
Performance in professional benchmarks
GPT-5.4 achieved 83.0% wins or draws against professionals in GDPval assessments, which covers 44 occupations and requires real products like accounting spreadsheets and presentations. Esse index surpasses the 70.9% of GPT-5.2. Executivos from companies like Mercor and Harvey highlight superiority in long-term deliverables and complex legal analysis.
The model improves spreadsheet creation and editing, with an average score of 87.3% in internal financial modeling tests, compared to 68.4% for its predecessor. Apresentações generated by GPT-5.4 were preferred in 68% of cases by human evaluators due to refined aesthetics and effective use of images.
Native computer usage capabilities
GPT-5.4 introduces native support for computer operation, allowing agents to execute complex flows across varied applications. Ele supports up to 1 million context tokens for planning and verification over long horizons. The model stands out in benchmarks such as OSWorld-Verified, with a success rate of 75.0%, and WebArena-Verified, reaching 67.3%.
Enhanced visual perception contributes to analysis of high-resolution documents and images. On the MMMU-Pro, accuracy reaches 81.2% without tools. Image input detail supports up to 2.56 million pixels, improving localization and click accuracy.
Advances in coding and efficiency
GPT-5.4 matches or surpasses GPT-5.3-Codex on SWE-Bench Pro, with lower latency in reasoning steps. /fast mode in Codex increases tokenization speed by up to 1.5 times. Tarefas complex front-ends generate more aesthetic and functional results.
An experimental skill called Dramaturgo Interativo allows visual debugging of web applications and Electron. Exemplos include theme park simulations created from simple prompts, with automated testing via Playwright.
Improvements in the use of tools and search
Searching for API tools reduces tokens needed across broad ecosystems, saving 47% on Atlas MCP tasks. Agentic tool calling increases accuracy in Toolathlon. Persistent web search improves on BrowseComp, with 82.7% accuracy.
GPT-5.4 Thinking describes its reasoning on long queries, allowing targeting without restarts. Ele maintains coherence in long streams and reduces hallucinations, with 33% fewer false statements and 18% fewer errors in complete responses.
Availability and pricing
The rollout occurs gradually on ChatGPT and Codex. In the API, GPT-5.4 is available as gpt-5.4 and version Pro as gpt-5.4-pro. Usuários Plus, Team and Pro access GPT-5.4 Thinking, replacing GPT-5.2 Thinking, which remains for three months. Planos Enterprise and Edu enable access via administrative settings.
The price reflects enhanced capabilities, but token efficiency pays off in many tasks. Batch Processamento and Flex cost half the standard, while priority doubles the price.






