On March 11, OpenAI quietly switched off its GPT-5.1 model line. Millions of active conversations were auto-migrated to newer equivalents.
GPT-5.1 Instant became GPT-5.3 Instant. GPT-5.1 Thinking became GPT-5.4 Thinking. GPT-5.1 Pro became GPT-5.4 Pro.
Six days earlier, GPT-5.4 had launched. It is described as the company’s most capable and efficient frontier model for professional work.
The model scored 83% on OpenAI’s GDPval benchmark for knowledge work tasks across 44 occupations. That places it at or above expert human level in most categories.
GPT-5.4 also took the lead on Mercor’s APEX-Agents benchmark for professional skills in law and finance.
Mercor CEO Brendan Foody said GPT-5.4 “excels at creating long-horizon deliverables such as slide decks, financial models, and legal analysis,” while being faster and cheaper than rivals.
GPT-5.4 is OpenAI’s first general-purpose model with native computer-use capabilities. It can control software directly using screenshots, mouse commands, and keyboard inputs.
On the OSWorld-Verified benchmark for computer control tasks, it achieved a 75% success rate — surpassing the human benchmark of 72.4%.
The model is also 33% less likely to make errors in individual factual claims compared to GPT-5.2, and 18% less likely to produce an error-containing full response.
ChatGPT’s context window for API users now supports 1 million tokens. That matches the context window offerings from Google and Anthropic.
The model release cadence is now measured in weeks. GPT-5.1 launched in January. GPT-5.3 in late February. GPT-5.4 on March 5.
For enterprise customers trying to evaluate, test, and deploy AI models in production, the pace is becoming its own kind of problem.