Friday, Jul 10 | --:--
Back to home

Models

Model Release United States

OpenAI Releases GPT-5.4 Focused on Agentic Workflows and Computer Use

OpenAI has launched GPT-5.4, its latest frontier model optimized for professional work with strong emphasis on agentic capabilities, coding, and native computer use. It introduces configurable reasoning effort and scores highly on benchmarks like SWE-bench and OSWorld.

Tech Insights Reporter 4 min read San Francisco

TLDR

OpenAI unveiled GPT-5.4 (along with Pro, mini, and nano variants) as its most capable model for agent-style tasks on March 5, 2026. It features native computer use APIs, a large context window (up to 1M tokens), and significant gains in coding and reasoning.

Key Capabilities

  • Configurable "reasoning effort" levels for balancing speed and depth.
  • Strong performance on computer use benchmarks (around 75% on OSWorld, slightly above human baseline of 72.4%).
  • Optimized for long-running agentic workflows and professional tasks involving spreadsheets, presentations, and documents.
  • Available in ChatGPT and via API, with Codex integration.
  • New "Tool Search" architecture for dynamic tool calling.

The release comes as OpenAI shifts resources to focus on developer and agent tools.

Industry Context

The launch follows closely on updates from other labs, compressing the timeline for frontier model improvements. Early benchmarks show it competitive with or leading in coding and agent tasks.

Why this story matters

The move toward explicit agentic and tool-use capabilities signals a broader industry shift from chat interfaces to systems that can take actions across software environments. This has major implications for automation in software development, enterprise workflows, and beyond.

It also underscores the intense competition, with labs rapidly iterating on practical capabilities rather than just raw scale.

Sources

  • OpenAI official announcement: "Introducing GPT-5.4" (openai.com/index/introducing-gpt-5-4/, March 5, 2026).
  • API documentation and pricing for GPT-5.4 variants.
  • Benchmark reports (SWE-bench, OSWorld comparisons) and TechCrunch coverage (March 5, 2026).

Featured Image Alt Text

GPT-5.4 interface showing agentic task execution and reasoning trace

Tags

OpenAI, GPT-5.4, Agents, Coding, Model Release

Scroll to continue reading