GPT‑5.4 boosts productivity
GPT‑5.4 just posted an 83% score on the GDPval benchmark — a 13-point improvement over its predecessor — and is being hailed as a major jump for knowledge-work productivity (nxcode.io). That leap matters for entrepreneurs: better research, synthesis and forecasting chops mean faster financial models, sharper market notes and more reliable signal generation for trading and personal finance workflows ( ).
OpenAI released GPT‑5.4 into ChatGPT, the API and Codex on March 5, 2026, and it is available in both a high‑performance "Pro" tier and a reasoning‑focused "Thinking" tier. (openai.com) The model supports up to a 1,000,000‑token context window and is the company’s first general model with native computer‑use capabilities that let agents view screens and execute multi‑step desktop workflows. (openai.com) OpenAI’s launch materials list new peaks on several industry tests, including SWE‑Bench Pro at 57.7%, OSWorld‑Verified at 75.0%, Toolathlon at 54.6%, and BrowseComp at 82.7% for the GPT‑5.4 configuration. (openai.com) OpenAI says GPT‑5.4 is its most token‑efficient reasoning model to date, using significantly fewer tokens than GPT‑5.2 to solve the same problems, a change OpenAI links to lower effective costs for long professional runs. (openai.com) TechCrunch reported OpenAI’s internal evaluations showing fewer factual‑claim errors and overall response errors versus prior models, and that GPT‑5.4 led Mercor’s APEX‑Agents benchmark designed to test law and finance workflows. (techcrunch.com) The product includes a new "Tool Search" system that looks up tool definitions on demand to reduce token overhead in tool‑rich environments, and OpenAI added a safety evaluation for chain‑of‑thought that showed reduced deceptive behavior in the Thinking mode. (techcrunch.com) OpenAI positioned GPT‑5.4 as combining the coding strengths of GPT‑5.3‑Codex with expanded reasoning and tool integration, while independent comparisons note GPT‑5.3‑Codex still holds advantages on some terminal‑centric coding tasks. (openai.com)
