OpenAI has unveiled GPT-5.4, an upgrade of its flagship artificial intelligence (AI) that introduces a transparent reasoning process and a tiered family of models designed to balance high-level professional work with cost-effective automation.
The release marks a move from black box responses to a more collaborative user experience. GPT-5.4 now features an integrated Thinking Mode, which uses an internal monologue to reason through complex prompts before generating text. To give users more control, the model produces a preamble during multi-step tasks, allowing humans to interrupt and redirect the AI if it veers off course.
The update also enhances ChatGPT’s Deep Research capabilities, enabling the system to scour the internet and synthesize reports from hundreds of documents while strictly adhering to the original prompt’s intent.
In internal benchmarks, GPT-5.4 surpassed Anthropic’s Claude 4.6 in mathematics and science, and matched or exceeded human professional output in 83% of knowledge-work tasks on the company’s GDPval scale.
Recognizing that not every task requires the massive compute power of a flagship model, OpenAI introduced a tiered lineup: GPT-5.4 Pro, a high-performance version for the most demanding logic and reasoning tasks; GPT-5.4 Mini, which runs twice as fast as its predecessor and is available to Free and Go users via the Thinking menu; and GPT-5.4 Nano, an API-exclusive model optimized for high-volume, low-cost tasks like data classification and extraction.
“The best model is often not the largest one,” OpenAI said in announcing the new model, “it’s the one that can respond quickly, use tools reliably, and still perform well.”
The agentic approach lets developers use a senior model (GPT-5.4) to plan tasks while delegating sub-tasks such as searching codebases or interpreting screenshots to the faster, cheaper Nano model.
Early adopters in specialized fields are already reporting a higher ceiling for accuracy. Niko Grupen, head of applied research at legal startup Harvey, noted that the model “sets a new bar for document-heavy legal work,” specifically citing its ability to maintain context across lengthy, complex contracts.
While efficiency has improved, the price for top-tier intelligence is rising. GPT-5.4 is priced at $2.50 per million input tokens, a jump from GPT-5.2’s $1.75. The Pro version commands a significant premium at $30 per million input tokens. However, the Nano model offers a budget-friendly alternative at just 20 cents per million tokens, aiming to make large-scale AI deployment more sustainable for startups.
The new ecosystem also integrates directly into productivity suites, with custom versions of ChatGPT now built into Excel and Google Sheets to automate financial modeling, earnings reports, and investment memos.
