On
OpenAI Unveils GPT-5.4: The Most Capable Model Yet, Built for Professional Work

OpenAI just dropped GPT-5.4, and it's being positioned as the company's most powerful and efficient model to date. What makes this release interesting is how OpenAI has bundled together advanced reasoning, coding prowess, and agent-like workflows—features previously scattered across specialized models—into a single, versatile AI system. The result? GPT-5.4 handles professional tasks like spreadsheet creation, document drafting, and presentation design with impressive competence.

GPT-5.4 Is Now the Strongest Model Available, Optimized for Professional Tasks

OpenAI claims GPT-5.4 tackles complex work with noticeably higher accuracy, efficiency, and speed compared to its predecessors.

The new model is available across ChatGPT, Codex, and via API for developers, in two distinct flavors. ChatGPT users get access to GPT-5.4 Thinking and GPT-5.4 Pro. Here's the clever bit: when you ask it something, the model first shows you its reasoning plan—how it intends to approach and solve your problem. You can then steer it in a different direction before it finishes generating the answer.

OpenAI also emphasizes that GPT-5.4 outperforms GPT-5.3 Instant when handling in-depth web research queries.

Another standout improvement: computer control. GPT-5.4 scored 75% on the OSWorld-Verified Computer Use benchmark—a significant leap. Plus, OpenAI is offering a preview with support for a massive 1 million token context window.

Benchmark performance is genuinely impressive. On GDPval, GPT-5.4 hit 83%—a new record. On the MMMU-ProGemini 3.1 Pro's 80.5%.

One persistent complaint about OpenAI's GPT models has been their struggle with creating visually polished documents and presentations. OpenAI says they've fixed that in GPT-5.4. The model now generates documents with better layouts, more diverse visual styles, and smarter use of image generation capabilities.

Beyond raw performance, OpenAI claims GPT-5.4 is their most token-efficient reasoning model yet.

Developers also get a new /fast mode in Codex. It speeds up token processing by up to 1.5x when running GPT-5.4 without sacrificing reasoning quality. The API version is also smarter about tool usage: instead of requiring the full tool definitions in your prompt upfront, GPT-5.4 accepts a compact tool list and searches for tools on demand.

For access, GPT-5.4 Thinking is now live for ChatGPT Plus, Team, and Pro subscribers—replacing GPT-5.2 Thinking. The older version will stick around for about three months. Meanwhile, GPT-5.4 Pro is exclusive to ChatGPT Pro users and Enterprise customers.

API pricing: gpt-5.4 costs $2.50 per million input tokens and $15 per million output tokens. The gpt-5.4-pro variant is pricier at $30 per million input tokens and $180 per million output tokens.

What's New in GPT-5.4

GPT-5.4 is OpenAI's latest flagship model, merging their best work in reasoning, coding, and computer interaction into one package.

It replaces GPT-5.2 Thinking in ChatGPT, is available through API and Codex, and comes with an experimental 1 million token context window in Codex. A Pro version is also included.

1 Million Token Context Window (Experimental in Codex)

The standard context window is 272k tokens, but Codex users can now configure GPT-5.4 to use up to 1 million tokens—matching competitors like Gemini 3 and Sonnet 4.6.

This expanded window handles long-form tasks where the model needs to plan, execute, and verify work across a much broader scope than previously possible.

Tool Search in the API

Tool search is a fresh API feature that loads tool definitions on-demand instead of bundling everything at once. Without it, massive tool ecosystems could add tens of thousands of tokens to every request. The efficiency gains are substantial.

Direct Computer Control

This is a big deal. GPT-5.4 is OpenAI's first general-purpose model with built-in computer interaction. It can interact with your desktop via screenshots, control your mouse and keyboard, and write Playwright code to automate browser tasks.

Better Spreadsheet and Presentation Creation

GPT-5.4 scores higher on spreadsheet modeling tasks, and evaluators prefer its presentation results compared to GPT-5.2. The main difference shows up in formatting and visual layout.

Fewer Hallucinations

GPT-5.4 is OpenAI's most accurate model to date. Individual statements are 33% less likely to be wrong than GPT-5.2, and complete answers are 18% less likely to contain any errors. These numbers come from anonymized prompts where users flagged actual mistakes.

Steering Capability

With long, complex queries, the new model sketches out its plan first—similar to Codex. You can then add guidance or adjust the direction if you're unhappy with its approach or changed your mind after submitting. What's interesting here is that steering has proven valuable for coding tasks, and GPT-5.4 brings this capability to other domains too.

Related Articles